Artificial intelligence (AI) is making rapid inroads into optometric and ophthalmic practices worldwide. Systems like RetinaLyze provide enhanced capabilities to assist with disease detection—offering faster analyses, automated screenings, and potentially more consistent patient outcomes. Yet, implementing AI in a clinical setting can raise important questions: Is this device properly regulated? How accurate is it for my patient population? Will it fit my workflow?

In a recent article published in Clinical and Experimental Optometry, Murphy et al. (2023) outlined a practical guide to help optometrists evaluate AI solutions. Drawing from their insights, here is a concise overview of how clinicians can appraise and integrate AI-based technologies such as RetinaLyze into everyday practice.

1. Check for Regulatory Approval

Why it matters:
Medical devices—AI-based or otherwise—must meet stringent standards to ensure safety and efficacy. Before introducing any AI system into your clinic, confirm that it has regulatory approval in your region. This helps guarantee that the device’s performance is validated by recognized authorities.

• Where to look:
  • In Australia, consult the Australian Register of Therapeutic Goods (ARTG).
  • In the United States, check the Food and Drug Administration (FDA).

Key takeaway:
If a product lacks regulatory clearance, it should not be used in patient care settings—even in prototypes—unless it is being studied under an approved clinical trial protocol.

‍

2. Understand the Intended Use

Why it matters:
AI systems are designed for specific tasks and may not work reliably if used “off-label.” For example, a device intended for measuring pupil size in contact lens fittings may not deliver dependable results if used to diagnose anisocoria. AI is not easily repurposed for tasks beyond its primary design.

• Check the official “intended use”:
  • Regulatory body listings (TGA, FDA, etc.) typically spell out the intended use of medical devices and AI software.

Key takeaway:
Use AI systems for their intended purpose. If you choose an off-label application, proceed with extreme caution: AI tools are optimized for particular datasets and tasks, and their performance may deteriorate outside of those parameters.

‍

3. Consider Your Clinical Workflow

Why it matters:
Sensitivity and specificity statistics can significantly impact the utility of an AI tool. A model with very high sensitivity but low specificity might flag too many false positives, creating unnecessary patient referrals and more work for clinicians.

• Ask yourself:
  • Will false positives or negatives be manageable in my practice?
  • How much time, manpower, and additional testing can we handle?

Key takeaway:
Choose an AI tool that aligns well with your practice’s objectives. For instance, if you need a first-line screening tool for diabetic retinopathy, a system with very high sensitivity (like RetinaLyze) might be ideal—provided you have a workflow in place to handle any false positives. RetinaLyze also provide a tele-medicine solution to handle these.

‍

4. Verify Representation in Test Data

Why it matters:
No two patient populations are identical. AI models trained primarily on one demographic may underperform or introduce bias when used on a different group (for example, minority communities). It’s important to ensure the AI has been validated in patient populations that resemble your own.

• Where to find this information:
  • Regulatory databases, product documentation, or published validation studies.
  • If data are not readily available, reach out to the manufacturer for more details.

Key takeaway:
Confirm your target population is represented in the AI’s test and training data. If no evidence of representative testing is available, use the tool cautiously and consider additional in-house validation or pilot testing.

‍

5. Look for Explainability (When It Matters)

Why it matters:
AI “black box” systems can be challenging to trust when you cannot see how they generate their results. However, not every clinical application demands full transparency—sometimes, it’s enough to know that the technology reliably screens for a condition without needing a detailed explanation of how it arrived at the conclusion.

• Examples of differing needs:
  • High-risk tasks (e.g., referable diabetic retinopathy diagnosis): In these cases, an AI that can provide “heatmaps” or annotated images allows a clinician to judge the validity of the system’s conclusions.
  • Lower-risk tasks (e.g., pre-fitting contact lens suitability): If the AI tool merely flags potential candidates for further assessment, it may not require the same level of detailed explanation. The clinician will conduct a thorough examination anyway, minimizing potential risk.

Key takeaway:
Match the level of needed AI transparency to the clinical risk. If the cost or risk of a misdiagnosis is high, opt for tools (like RetinaLyze screening solutions) that provide interpretable outputs.

‍

Conclusion

Artificial intelligence holds incredible promise for improving patient outcomes, increasing clinic efficiency, and aiding in the early detection of sight-threatening conditions. However, integrating AI tools such as RetinaLyze into clinical practice demands careful evaluation. By checking regulatory approvals, clarifying a system’s intended use, considering clinical workflows, ensuring adequate demographic representation in data, and balancing the need for explainable outputs, optometrists can confidently select the most appropriate AI solutions for their patients.

Interested in exploring how RetinaLyze can enhance your practice?
Feel free to get in touch with us for more information or a demonstration. By following the steps outlined by leading experts like Murphy et al., clinicians can harness the power of AI responsibly and deliver more comprehensive, efficient eye care to their patients.