PRESCRIPTION DIGITISATION USING BI-LSTM AND CTC-ENHANCED NEURAL NETWORKS

Garima Saroj, Ayushi Gautam, Bhavya Agrawal, Sarthak Gupta

Abstract: Handwriting is a way to convey an idea or information through written means. Over the decades, due to growing population density, doctors have been known for their often rushed and difficult-to-read handwriting. This readability issue of handwritten medical prescriptions, especially those created by clinicians, has an underlying potential to cause serious problems, ranging from confusion and delays in proper medication, to harmful side-effects. The aim of this research is to explore how deep learning techniques can help to reduce such risks by developing a model capable of reading and converting handwritten medical notes into digitised text. The model is based on a Region-based Convolutional Recurrent Neural Network (R-CRNN) and incorporates Connectionist Temporal Classification (CTC) loss which allows the model to recognise the handwritten characters, even in distorted or cursive writing. A combination of handwritten prescription samples and publicly available handwritten datasets were collected and annotated, preprocessed, and then the model training was performed. The research also aims to contribute towards the improvement of healthcare documentation and enhance the accessibility of medical records for people’s well-being and a sustainable future.

Keywords: Bidirectional Long-Short Term Memory (Bi-LSTM), Connectionist Temporal Classification (CTC) loss, Sequence modelling, Handwriting recognition, Optical character recognition (OCR), Deep learning

DOI: 10.24874/PES08.02A.018

Recieved:   Revised:   Accepted:   
UDC:

Reads: 2   

---