Rajeemanju/prescription-text-classifier

Prescription Text Classifier (DistilBERT)

A fine-tuned DistilBERT model that normalizes noisy prescription text (brand names, typos, doses, frequencies) to a standard generic medicine name.

Example: tab nappa 500mg 1+0+1 5 days → paracetamol

Model Details

• Developed by: Rajeemanju
• Model type: Transformer (DistilBERT) for sequence classification
• Language: English
• Base model: distilbert-base-uncased
• Task: Medicine name normalization (multi-class text classification)

Intended Use

For research and learning on medical text normalization, for example as a step in a prescription-processing pipeline.

Out-of-Scope Use

Not for real clinical or medical decisions. It has not been validated on real patient prescriptions.

Training Data

Custom dataset created by the author: [ ] samples, [ ] medicine classes, columns text and label. Text includes medicine name, dose, frequency, duration, and random typos. Split: 70% train, 15% validation, 15% test.

Training Procedure

1. Stage 1 (baseline): encoder frozen, only the classification head trained (4 epochs, learning rate 1e-3).
2. Stage 2 (fine-tuning): all layers unfrozen (6 epochs, learning rate 3e-5, weight decay 0.01, warmup 10%).

• Framework: Hugging Face Transformers (PyTorch), trained on Google Colab (T4 GPU)
• Max sequence length: 32

Results

Stage	Test Accuracy
Before fine-tuning (frozen encoder)	[ ]%
After fine-tuning	~98%

How to Use

from transformers import pipeline

clf = pipeline("text-classification", model="Rajeemanju/prescription-text-classifier")
print(clf("tab nappa 500mg 1+0+1 5 days"))

Limitations and Bias

• Covers only the medicines present in the training dataset.
• Trained on [synthetic / self-created] data, so performance on real-world or handwritten prescriptions may be lower.
• Text only. No handwriting or OCR.

Future Work

Larger real-world dataset, OCR for handwritten prescriptions, sequence-to-sequence normalization, more medicines.