Handwritten Clinical Chart Digitization & HITL Pipeline
Multimodal Ophthalmology Extraction with Click-to-Verify HITL — Vision-language pipeline digitizing handwritten clinical charts with dense cursive handwriting, tabular IOP logs, and an interactive click-to-verify human-in-the-loop workflow.
The Clinical Challenge in Ophthalmology
Ophthalmology charts present extreme digitization challenges: rapid cursive handwriting, proprietary shorthand, dense tabular IOP (intraocular pressure) logs, and physical lens (IOL) stickers attached to paper charts.
17-Model Benchmark & HITL Architecture
Benchmarked 17 frontier vision-language models for extraction accuracy vs. latency vs. cost, establishing an optimal pipeline running at just ~$0.012/page. Paired this with a progressive streaming verification interface where human reviewers click to verify low-confidence fields.
Current Milestones & Ongoing Work
Reduced scribe transcription time by 80–90%. Incorporating automated blank-page suppression and clinical range sanity checks to ensure 100% safety before EMR integration.