Privacy-Preserving OCR for Patient Documents

Abstract

Handwritten prescriptions and clinical notes remain common in many healthcare settings, yet their automatic processing is still highly challenging. Variability in handwriting styles, domain-specific medical terminology, and the frequent presence of abbreviations all contribute to high recognition difficulty. At the same time, recent progress in machine learning and OCR has led to powerful methods capable of learning complex patterns from data, but their behavior and limitations in clinical handwritten scenarios are still not fully understood. This project investigates how modern OCR techniques can be adapted and evaluated for handwritten medical documents, with a particular focus on prescriptions and short doctor’s notes. A key idea is to explore the incorporation of context—such as typical medication names, disease–treatment relationships, or common prescription structures—to help disambiguate unclear handwriting and improve recognition quality.