A CNN–BiLSTM–CTC Recognition System with Post-Processing Optimization for Handwritten English Prescription Digitization

Abstract: The digitization of handwritten medical prescriptions remains a critical challenge in healthcare informatics due to the prevalence of illegible handwriting, abbreviated terminology, and inconsistent writing styles. Errors in prescription interpretation contribute significantly to adverse drug events, medication misuse, and patient safety risks. This paper presents a robust offline handwritten prescription recognition framework based on a hybrid Convolutional Neural Network (CNN), Bidirectional Long Short-Term Memory (BiLSTM), and Connectionist Temporal Classification (CTC) architecture. The proposed system combines deep visual feature extraction using ResNet34 with sequential context modeling through BiLSTM layers and alignment-free transcription using CTC decoding. To improve clinical reliability, a post-processing module integrates fuzzy string matching and pharmaceutical contextual validation through Market Basket Analysis (MBA). The framework employs a modular segmentation strategy utilizing Single Shot Detector (SSD) based word localization and line-clustering heuristics to improve recognition quality. Experimental evaluation demonstrates a Character Error Rate (CER) of 8.5%, Character Recognition Rate (CRR) of 99.20%, and Word Recognition Rate (WRR) of 93.75% for pharmaceutical terms containing 7–10 characters. The proposed system achieves high recognition accuracy while maintaining computational efficiency, requiring substantially less memory than contemporary Vision Language Models (VLMs). The framework provides a practical, deployable solution for real-time prescription digitization and integration into healthcare information systems.

Keywords: Handwritten Text Recognition; Medical Prescription Recognition; CNN–BiLSTM–CTC; Deep Learning; Optical Character Recognition; Healthcare Informatics;