Residual U-Net Handwritten Prescription Recognition for Low-Resource Healthcare: Development and Validation Against Word and Character Error Targets

Medication errors remain a persistent patient-safety problem, and prescription illegibility is especially consequential where health systems continue to depend on handwritten medication orders. This study developed and validated a Residual U-Net deep learning model for handwritten prescription recognition, targeting a Word Error Rate (WER) below 15% and a Character Error Rate (CER) below 5% across varied handwriting conditions. A quantitative experimental design used the public Doctor's Handwritten Prescription BD dataset, comprising 4,680 segmented pharmaceutical-word images divided into 60% training, 20% validation and 20% testing sets. The reported model combined residual convolutional feature extraction, encoder-decoder recognition, transfer-learning principles, sequence modelling and automated hyperparameter search. Training proceeded for 50 epochs after Optuna-based preliminary optimisation. The best preliminary trial returned a reported CER objective of 0.3253 and selected a learning rate of 0.0005028, Adam optimisation, batch size 8, encoder LSTM hidden size 256 and decoder hidden size 512. Average training loss declined from 2.6637 at epoch 1 to 1.2816 by epoch 10 and 1.1845 by epoch 50. On held-out evaluation, the raw Residual U-Net achieved WER 8.5% and CER 2.1%, comfortably satisfying the pre-specified thresholds before lexicon-based correction. The WER was 6.5 percentage points below the maximum target and the CER was 2.9 percentage points below its maximum target. These findings support Residual U-Net-style feature extraction as a viable recognition baseline under the study conditions. However, the Bangladesh-origin dataset, synthetic contextual augmentation and absence of stratified WER/CER by handwriting style limit direct clinical generalisability to South Sudan; prospective testing on locally collected prescriptions remains necessary before deployment.