Project Spotlight: Optimising AI Scribe Configuration for Clinical Consultations

At the Centre of Digital Excellence (CoDE), we are committed to ensuring that the integration of AI technologies into clinical practice is not only safe, but effective and meaningful for both clinicians and patients. Our first project focuses on a critical foundation for successful deployment of AI scribe systems: how we configure and set up the technology to get the best results during live consultations.

The Challenge

AI medical scribes have the potential to transform healthcare by reducing administrative burdens and improving the quality of clinical documentation. However, their effectiveness can be highly sensitive to the physical and acoustic environment in which they operate.

This project explored three core questions:

  • How does microphone placement impact transcription accuracy?

  • What role does background noise play in the performance of AI scribes?

  • Can AI systems reliably distinguish relevant from irrelevant (or misleading) information during a consultation?

 

Methodology

Working within our simulated clinical environment at CoDE, we tested a commercial AI scribe across a series of controlled consultation scenarios. These were designed to mimic real-world conditions, varying:

  • The distance between microphone and clinician/patient

  • The presence and type of environmental background noise (e.g. toddler chatter, rain, construction)

  • The inclusion of additional “informational noise” such as speculative diagnoses or off-topic conversations

The resulting summaries were then compared against high-quality reference versions to assess the type and frequency of errors: omissions, hallucinations, and incorrect inclusions.

Key Findings

  • Microphone Distance Matters
    Error rates—particularly omissions—increased sharply when microphones were placed more than 2 metres away from the consultation. Integrated laptop microphones performed particularly poorly. Surprisingly, microphone cost had less impact than placement.

  • Background Noise Degrades Accuracy
    AI scribes were most affected by background speech (e.g. toddler chatter) and continuous ambient noise (e.g. heavy rain). Volume matters: when background noise matched or exceeded voice volume, error rates rose significantly.

  • Informational Noise Is Well Handled—But Not Perfectly
    AI scribes were generally good at ignoring irrelevant conversation and speculative patient diagnoses, especially when clearly separated in context. However, some medically important details were still omitted, requiring clinician vigilance.

Recommendations for Implementation

  • Use Dedicated Microphones at <1 Metre Distance
    Avoid relying on integrated laptop mics. Consider clip-on or desktop microphones placed near the clinician. For multi-zone consultations (e.g. couch and desk), consider dual mic setups.

  • Mitigate Background Noise
    Consultations should take place in acoustically stable environments. Where possible, avoid overlapping speech or persistent ambient noise.

  • Continue Clinician Oversight
    Even with good configuration, AI scribes are not infallible. Clinicians must continue to review and validate summaries, especially where the consultation includes complexity or ambiguity.

  • Support Change with Training
    Include AI scribe awareness and training in onboarding for clinical teams. Develop pathways for feedback and iterative refinement of local setups.

Why it Matters

This foundational work underlines CoDE’s commitment to evidence-led deployment of AI in clinical settings. By identifying practical, real-world factors that affect AI scribe performance, we’re helping clinicians, IT teams and developers get the most out of this promising technology, right from the very first step.