Modern OCR narrows gap on doctors’ handwriting

A new benchmark comparing modern optical character recognition (OCR) systems for medical prescriptions has shown that compact contemporary OCR models can decipher doctors' handwriting far better than traditional systems, though they still fall short of the accuracy required for autonomous clinical use. The April 2026 test pitted PP-OCRv5, GLM-OCR, Tesseract, and EasyOCR against each other on cropped handwritten prescription words.

GLM-OCR achieved the highest character-level performance with a Character Error Rate (CER) of 0.328, while PP-OCRv5 provided the lower Word Error Rate (WER) of 0.789. Despite the significant difference in error rates, exact-word recognition showed a sharper divide, with GLM-OCR correctly reproducing only 32.6% of tested words, compared to PP-OCRv5's 21.4%.

The results highlight a shift in OCR development, as PP-OCRv5, despite having only five million parameters, showed competitive word-level performance with GLM-OCR, a 900-million-parameter multimodal model. GLM-OCR employs a 400-million-parameter visual encoder and a 500-million-parameter language decoder, demonstrating the benefit of linguistic context for improved transcription.

However, generative OCR models like GLM-OCR also introduce new failure modes, such as producing formatting characters not present in the handwritten input. The experiment used the RxHandBD dataset containing 5,578 cropped word images and 1,559 unique text entries, which is limited to cropped individual words rather than full prescription pages.

Despite achieving a 32.6% exact-match score, a system making errors at this rate is not safe for clinical use, as roughly two out of three words would still be inaccurately reproduced. The benchmark was an independent technical test, not a clinical validation study, and did not evaluate patient outcomes or dispensing safety.