Week 8: Human-AI Teaming and Workflow Integration¶

Section 3: Application | Learning Outcome 3¶

Theme: Making it work: Integrating AI into clinical practice

Core Question: "How do we govern it?"

Learning Objectives¶

By the end of this session, students will be able to:

Design clinical workflows that optimise human-AI collaboration
Develop strategies to mitigate automation bias and alert fatigue
Create training programs for clinical AI users
Apply human factors principles to AI interface design evaluation

Content¶

8.1 Workflow Integration Principles¶

AI as Augmentation vs. Automation

Choosing the right model for different applications:

Approach	AI Role	Human Role	Best For
Augmentation	Provides additional information	Makes all decisions	Complex, uncertain situations
Automation	Performs routine tasks	Supervises, handles exceptions	Well-defined, repetitive tasks
Hybrid	Varies by situation	Varies by situation	Mixed complexity environments

Workflow Analysis

Understanding current state before introducing AI: - Task analysis: What do people actually do? - Information flow: What data moves where? - Decision points: Where are judgements made? - Failure modes: Where do things go wrong?

Integration Points

Where AI adds value vs. where it creates friction:

High-value integration points: - Information synthesis (gathering scattered data) - Pattern recognition (seeing what humans miss) - Consistency (reducing variability) - Speed (faster than human processing)

Potential friction points: - Decision interruption (breaking cognitive flow) - Additional steps (increasing task time) - Conflicting information (creating confusion) - Trust negotiation (deciding whether to follow AI)

The Problem of "One More Screen"

Cognitive load and attention are finite: - Each new system competes for attention - Information overload degrades performance - Integration vs. another standalone application - Screen time vs. patient time

8.2 Mitigating Human-AI Interaction Risks¶

Designing for Appropriate Trust

Neither over-reliance nor under-reliance: - Clear communication of AI certainty/uncertainty - Transparency about AI limitations - Feedback on AI accuracy over time - Calibration exercises

Alert Design Principles

Effective alerts are: - Infrequent enough to maintain attention - Salient enough to be noticed - Specific enough to guide action - Actionable (clear what to do)

Alert Fatigue Mitigation

Strategies: - Reduce alert volume (tune thresholds) - Tiered alerting (severity-based) - Contextual suppression (situation-appropriate) - Intelligent clustering (related alerts together) - Regular review of alert performance

Friction by Design

Sometimes making things harder improves safety: - Confirmation steps for high-risk actions - Requiring acknowledgment of AI uncertainty - Forcing consideration of alternatives - Preventing automatic acceptance

Feedback Loops

Helping clinicians calibrate their trust: - Show AI accuracy over time - Highlight cases where AI was right/wrong - Enable comparison with own accuracy - Celebrate appropriate overrides

8.3 Training and Competency¶

What Clinicians Need to Know

Essential knowledge for AI users: - What the AI does and doesn't do - How confident to be in outputs - When to trust and when to override - How to report concerns

Not usually needed: - Detailed technical architecture - Programming or data science skills - Mathematical foundations

Competency Frameworks

Components of AI competency: - Knowledge: Understanding what AI is and how it works (basic) - Skills: Using specific AI tools effectively - Attitudes: Appropriate trust, critical evaluation - Behaviours: Documentation, escalation, reporting

Training Approaches

Approach	Advantages	Disadvantages
Classroom/online	Efficient, consistent	Abstract, limited practice
Simulation	Realistic practice	Resource-intensive
Supervised practice	Real-world learning	Risk, variable quality
Peer learning	Accessible, contextual	May propagate misconceptions

Effective training typically combines approaches.

Simulation and Scenario-Based Training

Particularly valuable for: - Rare but critical AI failures - Override decision-making - Workflow integration - Team communication about AI

Ongoing Education

As systems evolve: - Updates when AI changes - Refresher training - Performance feedback - Learning from incidents

8.4 Change Management and Culture¶

Addressing Clinician Resistance

Distinguishing legitimate concerns from technophobia:

Legitimate concerns to address: - Patient safety - Professional autonomy - Workflow disruption - Accountability clarity - Job security

Misconceptions to correct: - AI will replace clinicians entirely - AI is always right - AI is completely untrustworthy - Nothing can be done to shape AI

Champions and Early Adopters

Leveraging informal influence: - Identify and support champions - Provide resources and time - Celebrate successes - Use peer influence

Managing Expectations

Neither overselling nor underselling: - Realistic benefits and timelines - Honest about limitations - Clear about learning curve - Commitment to address problems

Building a Learning Culture

Around AI use: - Psychological safety to report concerns - No blame for appropriate AI use decisions - Open discussion of errors - Continuous improvement mindset

Aeromedical Thread¶

Team Dynamics in Retrieval¶

AI Integration with Small Teams

Aeromedical teams are typically small (2-3 members): - Clear role allocation including AI - Communication about AI inputs - Shared awareness of AI status - Avoiding heads-down time

CRM Implications

Crew Resource Management with AI: - AI as "team member" - Speaking up about AI disagreement - Leadership when AI input conflicts - Situational awareness including AI

Different Crew Compositions

AI needs to work across: - Doctor-paramedic teams - Nurse-paramedic teams - Single-crew configurations - Varying experience levels

Training Requirements

For retrieval workforce: - Casual and rotational staff challenges - Maintaining currency - Simulation opportunities - Remote delivery options

Learning Activities¶

Pre-Class Preparation¶

Human Factors Reading
Complete assigned reading on human factors in healthcare AI
Note principles relevant to your practice
Workflow Observation
Observe and document a clinical workflow in your environment
Identify potential AI integration points
Note current pain points that AI might address

In-Class Activities¶

Workflow Redesign Workshop (Groups, 40 mins)
Groups receive current workflow and AI tool description
Design integrated workflow
Identify human factors considerations
Present and critique
Training Program Design (Pairs, 30 mins)
Design training program for AI tool implementation
Include: objectives, methods, assessment, ongoing support
Peer review and feedback
Interface Evaluation (Individual, 20 mins)
Evaluate AI interface using human factors checklist
Identify strengths and concerns
Suggest improvements

Post-Class Activities¶

Training Needs Analysis
Complete training needs analysis for selected AI implementation
Identify competency requirements
Design training approach

Indicative Resources¶

Required Reading¶

Sittig, D.F., & Singh, H. (2010). A new sociotechnical model for studying health information technology in complex adaptive healthcare systems. Quality and Safety in Health Care, 19(Suppl 3), i68-i74.

Session Summary¶

This week focused on the human and workflow dimensions of AI implementation:

Workflow integration requires understanding current practice before introducing AI
Human-AI interaction risks can be mitigated through thoughtful design
Training should focus on appropriate use, not technical mastery
Change management and culture are as important as technology
Aeromedical team dynamics create specific integration challenges

Next Week: We'll examine what happens after go-live—monitoring, maintenance, and incident response.