There’s a gap between AI hype and AI reality in healthcare. Vendors promise transformation. Conference presentations show impressive demos. But what’s actually deployable in an Australian hospital today, with TGA approval and reasonable implementation complexity?

I’ve evaluated dozens of clinical AI products over the past two years. Here are five application areas that I think are genuinely ready.

1. Chest X-Ray Triage and Detection

This is the most mature diagnostic AI category. Multiple products have TGA registration. Clinical evidence is substantial. Implementation pathways are well-understood.

What it does: Analyses chest X-rays to detect and flag critical findings—pneumonia, nodules, cardiomegaly, pneumothorax, and others. Most systems operate in advisory mode, providing a “second read” for radiologists.

Why it’s ready:

• Multiple TGA-registered products (including Annalise.ai, which is Australian-developed)
• Strong evidence base from large-scale clinical studies
• Clear workflow integration patterns
• Relatively straightforward PACS integration

Realistic expectations:

These systems are good at detection but not perfect. False positive rates vary by finding type. They work best as triage tools (prioritising urgent cases) rather than replacements for radiologist interpretation.

A reasonable target: reducing time-to-report for urgent findings by 30-50%. Not eliminating radiologist workload.

Implementation complexity: Medium. Expect 3-6 months from decision to go-live.

2. Diabetic Retinopathy Screening

AI for diabetic retinopathy screening has been in clinical use internationally for years. Australian adoption is growing.

What it does: Analyses retinal images to detect diabetic retinopathy, often in primary care or screening settings where ophthalmology access is limited.

Why it’s ready:

• FDA-approved products with Australian availability
• MBS-aligned screening pathways
• Strong evidence for screening accuracy comparable to specialists
• Addresses a genuine access problem (not enough ophthalmologists, especially in regional areas)

Realistic expectations:

These systems work best for screening—identifying who needs referral to a specialist. They’re not intended for diagnosis or treatment decisions.

The biggest impact is in settings that currently have no retinal screening because specialist access is limited.

Implementation complexity: Low to medium. Many systems are designed for deployment in primary care with minimal technical infrastructure.

3. Clinical Documentation and Coding Support

This isn’t diagnostic AI, but it’s highly practical. AI systems that help with clinical documentation and coding accuracy are delivering value today.

What it does: Analyses clinical documentation to suggest appropriate diagnosis codes, identify documentation gaps, and improve coding completeness.

Why it’s ready:

• Clear regulatory pathway (administrative AI, not medical device)
• Immediate ROI through improved coding accuracy and revenue capture
• Doesn’t require clinical workflow changes
• Lower risk profile than diagnostic applications

Realistic expectations:

Good systems improve coding accuracy by 10-20% and reduce documentation rework. They work as productivity tools for clinical coders and HIM professionals, not replacements.

They also help identify documentation quality issues that clinicians can address, which has indirect clinical benefits.

Implementation complexity: Low. Primarily a software deployment, not a clinical transformation project.

4. Sepsis Early Warning

AI-powered early warning systems for sepsis are showing promising results in real-world deployments.

What it does: Monitors patient vital signs and pathology results to identify patients at risk of sepsis before clinical deterioration becomes obvious.

Why it’s ready:

• Several systems have Australian deployments underway
• Sepsis is a high-value target (major cause of preventable mortality)
• Evidence from US deployments is encouraging
• TGA pathways for clinical decision support are clearer now

Realistic expectations:

The best systems can identify sepsis risk 4-6 hours earlier than traditional criteria. That time window matters—early intervention significantly improves outcomes.

But these systems need to be embedded in response pathways. An alert that doesn’t trigger appropriate clinical action has no value.

Implementation complexity: Medium to high. The AI is the easy part. Building response protocols and ensuring nursing staff act on alerts is the hard part.

5. Pathology Image Analysis (Selected Applications)

AI for pathology is progressing, though adoption is earlier than radiology. Selected applications are deployment-ready.

What it does: Analyses histopathology slides for specific findings—breast cancer biomarkers, prostate cancer grading, and others.

Why it’s ready (for selected applications):

• TGA-registered products for specific use cases
• Addresses pathologist workforce shortages
• Strong evidence for specific applications (Ki-67 scoring, HER2 analysis)
• Can improve consistency in subjective assessments

Realistic expectations:

This isn’t “AI reads pathology slides.” It’s AI assisting with specific quantitative tasks that are time-consuming and sometimes subjective.

Start with narrow applications where evidence is strongest. Don’t try to implement “general pathology AI”—it doesn’t exist yet.

Implementation complexity: Medium to high. Requires pathologist buy-in and often laboratory information system integration.

What’s Not Ready Yet

For balance, here are areas that get a lot of attention but aren’t ready for broad deployment:

Natural language processing for clinical notes. Promising but accuracy isn’t sufficient for clinical decision-making. Good for administrative applications, not clinical ones.

AI for radiology beyond detection. Systems that provide diagnoses or treatment recommendations (not just detecting findings) face higher regulatory hurdles.

Autonomous clinical decision systems. AI that makes decisions without human oversight isn’t approved and shouldn’t be.

General-purpose clinical AI. The dream of AI that handles diverse clinical scenarios remains distant. Current AI is narrow.

Choosing Where to Start

If you’re prioritising AI investments, I’d suggest these criteria:

Regulatory clarity. Is there a clear TGA pathway? Are products already registered?

Evidence strength. Is there published clinical evidence, not just vendor claims?

Workflow fit. Does it address a genuine workflow problem your clinicians recognise?

Implementation feasibility. Do you have the infrastructure, expertise, and clinical engagement to succeed?

Risk profile. Can you start with lower-risk applications to build capability before tackling higher-risk ones?

The five applications I’ve described score reasonably well on all these criteria. That’s why I think they’re ready.

Start with one. Get it working. Learn from the experience. Then expand.

---

Dr. Rebecca Liu is a health informatics specialist and former Chief Clinical Information Officer. She advises healthcare organisations on clinical AI strategy and implementation.