SimpleScribeVA v1.1.0 (transcription model included)

SimpleScribeVA v1.1.0

Warning: This offline virtual medical scribe, intended to work with the secure, HIPAA-compliant LLM of your choice, remains unapproved by the VA. I'm releasing it to allow for testing and feedback.

Background: Virtual medical scribes are rapidly becoming ubiquitous across health care, with both significant potential benefit as well as risk. The benefit of a VA-specific scribe is that it would allow for customization of prompts for professions that are currently underserved in the market (e.g. social workers and nurses), improved focus on Veteran-specific issues and programs like HUD-VASH, and eventual integration with VISTA/CPRS. Most importantly, careful internal testing would allow for rapid feedback and correction on any issues of concern that might arise.

Instructions: First make sure you have downloaded Python 3.13, and that during installation you checked the box to have it added to the Path for all users. Second, download this zip to your OneDrive desktop on a Windows computer and double-click to decompress. Double click on Setup.bat. This may take several minutes to finish, and you may need to repeat the Setup.bat process on new computers. Then double-click StartSimpleScribeVA.bat any time you'd like to get started. A browser will open to http://127.0.0.1:5000 and in a few moments you should see the SimpleScribeVA interface.

CHANGES IN THIS RELEASE

Improved Accuracy: This release uses a larger, more accurate Whisper transcription model (ggml-medium.en.bin as opposed to ggml-small.en.bin), without a significant delay in transcription time. Transcription remains entirely on your computer.

More User Feedback: This releases asks for JSON instead of TXT files from Whisper, and uses the metadata to provide confidence scores for tokens (words or syllables). Those below a certain confidence level are highlighted in the transcription box, so the user can quickly fix glaring issues right in the transcript box before copying to the clipboard. The transcript also strips out some disfluencies like um or ah, and periods of silence.

Default Prompts Improved: The default prompts now specify that you're caring for Veterans to allow for Veteran specific services like VASH to be incorporated more easily. More importantly, the prompts now ask GPT to slow down and take the additional sequential steps of attempting to 1) correct obvious transcription errors based on context 2) categorize statements (e.g. S, O, A, P) and 3) ask the user to approve the corrected, categorized statements before creating the note. This pre-categorization reduces hallucinations in the final product, and the incorporation of an additional user check improves overall reliability.