AI Clinical Notes

• Reduce admin time for allied health professionals.
• Strengthen adoption of VALD Hub as the centre of the clinical workflow.
• Close the competitive gap in AI scribing within allied health.
• Lay the groundwork for future integrations; first into patient timelines, then into case management for tracking patient journeys over time.

VALD’s unique position provided a competitive edge over standalone AI scribe tools:

• Largest normative dataset for physical health tests.
• Existing clinic integrations across measurement hardware and software.
• Exercise program creation already embedded in the platform.
• Test results from VALD’s measurement apps could be directly cited in clinical notes, eliminating redundancy.
• Historical data access practitioners could generate notes informed by the entire patient record.

Together, these factors can lead to a more robust note with less manual work required from the practitioner.

I spent time as a patient and observer understanding how physios conduct their sessions both with and without AI tools. Some key findings from mapping out the workflow included:

• Practitioners rarely revisited transcripts. They valued the note, not the raw conversation. The transcript would be viewed only when the note quality or content didn’t meet or exceed expectations.
• Intake assessments and follow-up sessions both required quick access to patient data.
• Free text was only used for shorthand or instructions. Transcribing the session and generating a note from the transcription provided the most value.
• AI tools were seen as promising but prone to uncertainty, trust required clear feedback and predictable states.

The following principles guided the solution:

• Speed & versatility: Reduce time behind screens, support flexible practitioner habits.
• Trust: Feedback loops for recording and note generation to reduce uncertainty.
• Clarity: Surface only essential actions; keep noise low. Provide clear direction with one primary action in any one state.
• Consistency: Align actions with practitioner expectations (e.g., start/stop recording).
• Focus on the note: Transcripts and free text were secondary; the generated note was the core outcome.

Using Cursor, I built a functional prototype integrating OpenAI’s Whisper and GPT models. This enabled us to test real-world recording, transcription, and note generation flows in a real environment and gain valuable insights.

Key insights from prototype testing:

• Practitioners often needed to add more context to the recording (i.e ‘resume recording’) once the patient had left to add clinical reasoning and definition to the note that may have been left out of the conversation with the patient.
• Resume recording was critical for multi-part sessions.
• Profile selection was straightforward; however, note template selection was often overlooked — likely a training issue rather than a UI one.
• Practitioners wanted confidence that recording was active and saving properly; reassurance states were essential.

We explored multiple interaction models:

• Recording flows: Start/stop vs. start/ generate. Practitioners preferred simple start/stop with the option to resume later, which felt more natural and aligned with the mental model people were used to with video/ voice recording.
• Note generation: Shifting from a manual “generate note” button to automatic generation upon stopping a recording simplified the experience and reduced confusion. It also enabled a more fluid state change
• Navigation: Tested state-driven pages vs. tabbed navigation. State-driven flows proved clearer for practitioners under time pressure. They also helped users focus on the primary task.
• Transcript handling: Deprioritised as it was not valuable in 99% of cases.
• Note editor: Balanced structured sections (e.g., SOAP) with open text for flexibility. We didn’t want to add up too many guard rails with fixed fields as thai would limit customisation and add complications with custom prompts in the future.

Through testing, we shifted the mental model:

• From “generate a note” → to “stop/ finish/ end recording = generate note”.
• This aligned better with how practitioners thought about recording their sessions.
• It also reassured the practitioner that the recording had stopped.
• The decision to allow users to navigate away from the note anywhere in the platform. This led to ‘Stop recording’ aligning better with the intended action when a live recording is in progress when on another page.
• Navigating away was important to enable users to view profile data such as previous tests run or intake assessment reports.
• Allowed navigation away from the session while continuing to record.