Towards Modeling Situational Awareness Through Visual Attention in Clinical Simulations

This study utilizes Transition Network Analysis on eye-tracking data from 40 clinicians in VR-based cardiac arrest simulations to demonstrate how visual attention dynamically redistributes across roles and scenario phases, offering a novel framework for modeling team situational awareness and enhancing acute care training.

The Gist

Imagine a high-stakes emergency room, but instead of a real hospital, it's a hyper-realistic video game (Virtual Reality). In this game, a team of four doctors has to save a patient whose heart has stopped. It's chaotic, fast, and terrifying.

The big question the researchers asked was: How does a team "know" what's happening around them without talking constantly? In psychology, this is called Situational Awareness (SA). It's the team's shared mental map of the crisis.

Usually, to study this, researchers have to watch videos of the doctors and manually code every time they look at a monitor or a person. It's slow, boring, and misses the split-second flow of information.

So, this team of researchers tried something new. They used eye-tracking glasses to see exactly where the doctors were looking, and then they used a math tool called Transition Network Analysis (TNA).

Here is the simple breakdown of what they found, using some everyday analogies:

1. The "Gaze Highway" (The Method)

Think of the doctors' eyes as cars driving on a highway.

• The Stops (AOIs): The "stops" on this highway are specific things in the room: The Patient, the Heart Monitor, the Defibrillator (shock machine), the CPR machine, and the other team members.
• The Traffic Flow: The researchers didn't just count how many times a doctor looked at the monitor. They mapped the traffic flow. Did the doctor look at the monitor, then the patient, then the shock machine? Or did they stare at the monitor for a long time?
• The Map: They turned these eye movements into a network map. If a doctor frequently looks back and forth between the "CPR Machine" and the "Patient," that creates a strong, thick line on their map.

2. The Two Main Metrics (The "Vibe" of the Team)

To understand the traffic, they measured two things:

• Entropy (The "Scatter" Factor): This measures how wild or organized the eye movements are.
  • High Entropy: The eyes are jumping everywhere, like a squirrel looking for nuts. This means the person is gathering lots of different clues.
  • Low Entropy: The eyes are focused on a few specific spots. This means the person is in "tunnel vision" mode, focusing deeply on one task.
• Self-Loop Rate (The "Stare" Factor): This measures how often a doctor looks at the same thing twice in a row without looking away.
  • High Self-Loop: They are staring intensely at one thing (like a driver staring at a pothole). This is good for doing a specific job but bad if you miss what's happening around you.

3. The Results: Different Roles, Different "Driving Styles"

The study found that the four roles on the team drove their "eye highways" very differently, and they changed their driving style as the emergency got worse.

The CPR Doctor (The "Doer")

• Role: They are the ones physically pushing on the chest.
• Early Game: At the start, they looked around a lot (High Entropy). They were checking the patient, the team, and the machines to figure out the plan.
• Late Game: Once they started pushing on the chest, their vision narrowed. They stopped looking at the team and the other machines. They stared almost exclusively at the patient and their own hands.
• The Analogy: Think of a race car driver. At the start of the race, they look at the whole track. But when they are cornering at 200 mph, they focus only on the apex of the turn. They have to "tunnel" to do the job right.

The Team Leader (The "Conductor")

• Role: They don't touch the patient; they watch the whole room and give orders.
• Early Game: They were looking at everything, trying to get the big picture.
• Late Game: As the CPR doctor got "tunnel vision," the Team Leader's eyes did the opposite. They started staring more at the heart monitor and the other team members.
• The Analogy: Think of a conductor of an orchestra. When the violinist (CPR) is playing a fast, intense solo, the conductor stops looking at the violinist and focuses on the rest of the orchestra to make sure everyone else is in sync. The Leader picks up the "scanning" job that the Doer dropped.

The Defibrillator Doctor (The "Specialist")

• Role: They manage the shock machine.
• Behavior: They tended to stare at their machine and the patient's heart rhythm. They had the highest "Stare Factor" (Self-Loop). They were like a surgeon focused entirely on their specific tool, not looking around much.

4. The Big Takeaway: The "Handoff" of Awareness

The most cool part of the study is that the team didn't just "share" awareness equally. Instead, they dynamically redistributed it.

• The Problem: If everyone tried to look at everything, they would get overwhelmed.
• The Solution: As the crisis got intense, the "Doers" (CPR) narrowed their focus to survive the immediate task. The "Leaders" (TeamLead) automatically expanded their focus to watch over the whole team.
• The Metaphor: Imagine a relay race. When one runner (CPR) is sprinting, they can't look at the finish line or the other runners; they just look at the track. The person waiting for the baton (TeamLead) has to watch the runner and the track to know when to start running. The "awareness" of the race is passed from one person to another depending on who is sprinting.

Why Does This Matter?

This study proves that Situational Awareness isn't a static thing you just "have." It's a living, breathing system that shifts like a dance.

By using this "Eye Traffic Map" (TNA), trainers can now see exactly when a team is breaking down.

• Example: If the CPR doctor is staring at the wall (too much tunnel vision) AND the Team Leader is also staring at the wall (not watching the team), the system is broken.
• Future: In the future, VR training could give real-time feedback: "Hey, your focus is too narrow, look up!" or "You're not watching the team, check on your partner!"

In short: The paper shows that in a crisis, a great team doesn't all look at the same thing. They specialize. The "Doers" focus on the task, and the "Leaders" focus on the big picture, and they swap these mental jobs seamlessly as the situation changes.

Technical Summary

Here is a detailed technical summary of the paper "Towards Modeling Situational Awareness Through Visual Attention in Clinical Simulations."

1. Problem Statement

Situational Awareness (SA) is critical for team performance in time-critical clinical environments (e.g., cardiac arrest), yet it is difficult to characterize due to its dynamic, distributed, and multimodal nature.

• Limitations of Current Methods: Traditional observational approaches (behavioral markers, coding schemes) are labor-intensive, subjective, and fail to capture the real-time, interdependent dynamics of team processes.
• Limitations of Existing Gaze Metrics: Standard eye-tracking metrics (fixation counts, dwell times, pairwise transitions) fail to capture the temporal structure and sequential reorganization of attention as task demands evolve across multiple team members.
• Research Gap: There is a lack of scalable, objective analytical approaches that can model how visual attention flows and adapts across different clinical roles and scenario phases to support collaborative problem-solving.

2. Methodology

The study employs Transition Network Analysis (TNA), a graph-based approach, to model visual attention as a temporal network of gaze transitions between clinically meaningful Areas of Interest (AOIs).

Study Context & Data Collection

• Setting: Multiperson Virtual Reality (VR) cardiac arrest simulations using HP Reverb G2 Omnicept headsets.
• Participants: 40 clinicians (paramedics, physicians, residents) divided into 10 teams of four.
• Roles: Standardized roles based on American Heart Association guidelines: TeamLead, Airway, CPR, Defib.
• Scenario: A STEMI cardiac arrest scenario progressing from initial assessment (Stage 1) to peak intervention/return of circulation (Stage 5).
• Data Processing:
  • Raw data: 55,670 fixations and 22,660 saccades.
  • Preprocessing: Adjacent fixations on the same object (gap ≤300 ms) were merged.
  • Final Dataset: 20,628 merged fixations and 20,526 AOI transitions.

Analytical Framework

1. AOI Definition: Seven specific Areas of Interest were defined (e.g., Equipment-Airway, Patient Vitals Monitor, Other Team Members, Patient).
2. Sequence Construction: Eye-tracking data was transformed into role-specific AOI fixation sequences ordered by timestamp.
3. Transition Matrix Estimation:
   • A $7 \times 7$transition count matrix ($C$) was computed for each participant-role.
   • Laplace Smoothing ($\alpha=0.5$) was applied to stabilize probabilities for sparse data.
   • Row normalization produced a row-stochastic transition probability matrix ($P$).
4. TNA Metrics Extraction:
   • Entropy ($H$): Calculated over non-self transitions ($i \neq j$) to measure the unpredictability/dispersion of attention. High entropy indicates diverse, exploratory scanning; low entropy indicates structured, predictable routing.
   • Self-Loop Rate: Calculated as the weighted proportion of transitions where the gaze remains on the same AOI ($P_{ii}$). High rates indicate attentional tunneling (narrowing focus on specific cues); low rates indicate broader, distributed monitoring.
5. Qualitative Analysis: Visual inspection of dyadic (two-node) and triadic (three-node) motifs to identify recurring attentional routines.

3. Key Contributions

• Novel Application of TNA in Healthcare: First application of Transition Network Analysis to model visual attention in multiperson VR-based clinical simulations, moving beyond static gaze metrics to dynamic network structures.
• Decoupling Metrics: Explicitly separates Entropy (exploratory scanning) from Self-Loop Rate (tunneling/sustained focus), providing a nuanced view of attentional structure that traditional metrics miss.
• Role-Specific & Stage-Specific Modeling: Demonstrates how attentional networks are not static but dynamically reorganize based on clinical role and the progression of the medical scenario.

4. Key Results

A. Role-Based Differences (Cross-Sectional)

• CPR Role: Exhibited the highest entropy and relatively lower self-loop rates. This suggests flexible, multi-cue monitoring necessary for executing chest compressions while tracking immediate patient response.
• Defib Role: Showed the highest self-loop rates, indicating "device-centric" attentional tunneling focused on the defibrillator unit and vitals.
• TeamLead & Airway: Occupied intermediate positions but showed more distributed networks. TeamLeads specifically bridged patient data, monitors, and team dynamics, reflecting a "bird's-eye view" consistent with leadership.
• Statistical Significance: Significant differences were found in both Entropy ($p \approx 0.0015$) and Self-loop rate ($p \approx 0.0016$) across roles.

B. Stage-Based Shifts (Longitudinal)

• CPR Role Evolution:
  • Stage 1 (Initial): Broad exploratory scanning across multiple AOIs (Defib, Patient, Team).
  • Stage 5 (Intervention): Network centralized around Patient and Equipment-CPR. Self-loops on execution-critical tasks increased, while peripheral scanning decreased. This represents an adaptive narrowing of focus to maintain output quality under high motor load.
• TeamLead Role Evolution:
  • Stage 1: Focused on synthesizing information (triadic motifs involving monitors, equipment, and team).
  • Stage 5: Shifted to monitoring-centric oversight. Self-loops on Patient Vitals Monitor and Other Team Members increased significantly ($0.61 \to 0.82$and$0.67 \to 0.74$).
• Functional Differentiation: As the CPR role's attentional bandwidth contracted for execution, the TeamLead role expanded its monitoring scope to maintain global team awareness. This supports distributed cognition theories, showing SA is dynamically redistributed rather than uniformly shared.

5. Significance & Implications

• Theoretical Advancement: Validates that SA in acute care is encoded in the structure and dynamics of visual attention networks. It confirms that SA is a dynamic, compensatory process where cognitive responsibilities are flexibly partitioned among team members.
• Practical Application:
  • Training Interventions: TNA can identify phase-specific vulnerabilities (e.g., a TeamLead failing to shift to monitoring mode during peak intervention).
  • Real-time Feedback: The metrics (entropy, self-loop) can serve as real-time indicators of team coordination health in VR training.
  • Objective Assessment: Provides a scalable, objective alternative to labor-intensive human coding for assessing team performance.
• Future Directions: The authors plan to integrate multimodal data (speech, physiological signals) and focus on event-centered analysis to capture attention reorganization around pivotal clinical moments.

Conclusion

The study successfully demonstrates that Transition Network Analysis is a powerful lens for mapping the functional differentiation of team cognition. It reveals that effective situational awareness in cardiac arrest is not a static state but a dynamically rebalanced system where roles adapt their visual attention strategies (from exploratory to tunneling, or vice versa) to meet evolving clinical demands.