Week 10: Current AI Landscape in Healthcare¶

Section 4: Synthesis | Learning Outcome 4¶

Theme: State of play: What's actually working (and what isn't)

Core Question: "What comes next?"

Learning Objectives¶

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

Survey the current state of AI deployment in Australian healthcare
Analyse international exemplars and their applicability to Australian context
Evaluate the evidence base for deployed healthcare AI systems
Identify gaps between AI hype and operational reality

Content¶

10.1 AI in Australian Healthcare Today¶

Deployed Systems

Current operational AI in Australian healthcare:

Medical Imaging - Radiology AI (chest X-ray, mammography, CT) - Ophthalmology screening (diabetic retinopathy) - Pathology AI (emerging) - Dermatology AI (limited deployment)

Clinical Decision Support - Sepsis prediction/alerting - Deterioration early warning - Drug interaction checking - Diagnostic support tools

Administrative - Clinical coding assistance - Scheduling optimisation - Documentation support - Revenue cycle applications

Australian Success Stories

Examples of effective AI implementation: - [Case studies to be developed with current examples] - Key success factors - Transferable lessons

Cautionary Tales

Examples where AI hasn't delivered: - [Case studies to be developed] - What went wrong - Lessons learned

The Vendor Landscape

Who's selling AI to Australian health services: - International vendors localising products - Australian-developed solutions - Academic/research spinoffs - Startup ecosystem - Procurement and evaluation challenges

10.2 International Perspectives¶

UK NHS AI Lab

National coordination approach: - Centralised evaluation and guidance - National implementation support - Evidence generation programs - Regulatory alignment

Lessons for Australia: - Benefits of coordination - Resource requirements - Transferable frameworks

US Health System Implementations

Scale and fragmentation: - Large health system deployments - Vendor-driven market - Variable evidence standards - Rapid commercialisation

Lessons for Australia: - Scale opportunities and risks - Evidence generation models - Regulatory differences

European Approaches

Regulation-led development: - EU AI Act implications - CE marking requirements - Privacy (GDPR) integration - Ethical frameworks

Lessons for Australia: - Regulatory alignment opportunities - Ethical framework development - Cross-border considerations

Emerging Economy Innovations

Necessity-driven AI: - Resource-constrained solutions - Task-shifting applications - Mobile-first approaches - Scalability lessons

Lessons for Australia: - Rural and remote applications - Efficient deployment models - Appropriate technology

10.3 Evidence and Outcomes¶

What Does the Evidence Actually Show?

Systematic review findings: - Many promising results in development - Fewer rigorous clinical trials - Limited real-world outcome data - Publication bias concerns

The Gap Between Trial and Reality

Why real-world performance differs: - Controlled conditions vs. messy reality - Selected populations vs. all comers - Research support vs. routine care - Novelty effects and attention

Measuring AI Impact

Dimensions of impact: - Clinical outcomes (mortality, morbidity) - Process outcomes (efficiency, timeliness) - Economic outcomes (cost, resource use) - Experience outcomes (clinician, patient)

Measurement challenges: - Attribution (was it the AI?) - Timeframes (when do benefits appear?) - Unintended consequences - System-level effects

Publication Bias and Missing Studies

What we don't know: - Failed implementations rarely published - Negative results underreported - Vendor-sponsored research biases - The graveyard of abandoned projects

10.4 Aeromedical and Emergency Services AI¶

International Retrieval Service Innovations

Dispatch and Tasking - AI-assisted triage - Resource allocation optimisation - Demand prediction

Clinical Decision Support - Diagnostic assistance - Treatment recommendations - Risk stratification

Operational Support - Flight planning - Weather integration - Crew scheduling

Emergency Dispatch AI

Current applications: - Call prioritisation - Protocol adherence support - Resource recommendations - Quality assurance

Evidence and concerns: - Performance data - Equity considerations - Implementation challenges

In-Field Diagnostic Support

Emerging capabilities: - Point-of-care ultrasound AI - ECG interpretation - Vital sign pattern recognition - Symptom checkers for field use

Search and Rescue Applications

AI applications in SAR: - Search pattern optimisation - Survivor detection - Resource coordination - Terrain and weather analysis

Aeromedical Thread¶

Deep Dive: International Aeromedical AI¶

Detailed examination of AI currently deployed or trialled in aeromedical services globally:

European Services - Scandinavian HEMS innovations - UK air ambulance technology - Swiss Rega systems

North American Services - US flight programs - Canadian remote area services - Integration approaches

Australian/NZ Developments - Current state assessment - Trials and pilots - Strategic directions

Critical Analysis

For each exemplar: - What's actually deployed vs. aspirational? - What evidence exists? - What would transfer to Australian context? - What barriers exist?

Learning Activities¶

Pre-Class Preparation¶

Environmental Scan
Identify AI currently deployed or considered in your health service
Assess current state: operational, pilot, evaluation, aspirational
International Case Study
Research one international AI implementation in detail
Prepare brief summary for class sharing

In-Class Activities¶

Evidence Appraisal Workshop (Groups, 40 mins)
Groups critically appraise published AI implementation studies
Assess strength of evidence
Identify gaps and limitations
Report key findings
Guest Presentations (If available, 30 mins)
Presentations from health services with deployed AI
Q&A and discussion
Lessons learned
Gap Analysis Exercise (Pairs, 20 mins)
Compare international exemplars to Australian context
Identify transferable elements
Note barriers and adaptations needed

Post-Class Activities¶

Current State Analysis
Complete current state analysis section for capstone project
Document relevant deployed systems and evidence base

Practical Exercise 10: Clinical LLM Experimentation (Colab)¶

Objective¶

Systematically evaluate large language model capabilities and limitations for clinical applications.

Part A: Setting Up LLM Access¶

import anthropic

client = anthropic.Anthropic(api_key="PROVIDED_FOR_COURSE")

def query_llm(prompt, system_prompt="You are a helpful clinical assistant."):
    response = client.messages.create(
        model="claude-sonnet-4-20250514",
        max_tokens=1024,
        system=system_prompt,
        messages=[{"role": "user", "content": prompt}]
    )
    return response.content[0].text

Part B: Clinical Knowledge Assessment¶

clinical_questions = [
    "What are the diagnostic criteria for sepsis according to Sepsis-3?",
    "Describe the management of a tension pneumothorax in the pre-hospital setting.",
    "What are the contraindications to thrombolysis in acute ischaemic stroke?",
]

for question in clinical_questions:
    print(f"Q: {question}\n")
    print(f"A: {query_llm(question)}\n")

Part C: Probing for Limitations¶

# Test for fabrication
response = query_llm("What did the RETRIEVAL-2 trial show about adrenaline dosing?")
# Note: This is a fictional trial

# Test for temporal limits
response = query_llm("What are the latest 2024 ANZCOR guidelines?")

# Test for Australian-specific knowledge
response = query_llm("What PBS restrictions apply to apixaban in Australia?")

Deliverable¶

Completed evaluation notebook with structured assessment of LLM capabilities and limitations for three clinical use cases.

Indicative Resources¶

Required Reading¶

Selected peer-reviewed AI implementation studies (provided)

Session Summary¶

This week surveyed the current state of healthcare AI:

AI is deployed in Australian healthcare, but scale and maturity vary
International exemplars offer lessons but require contextual adaptation
The evidence base for healthcare AI has significant gaps
Real-world performance often differs from published results
Aeromedical services internationally are exploring AI applications

Next Week: We'll look forward—examining emerging technologies and trends that will shape healthcare AI.