DistilBERT fine-tuned to classify patient queries into 9 clinical intent categories for NHS triage routing. Part of a project demonstrating how a model that scored 100% on a standard train/test split turned out to be only 18% reliable on independent, realistic patient phrasing — and what it took (LIME diagnosis + targeted LLM-generated augmentation across 12 iterations) to reach 92.3% on a leakage-checked stress test.
causes · exams and tests · information · inheritance · precautions · side effects · symptoms · treatment and outlook · when to contact a medical professional
1from transformers import AutoTokenizer, AutoModelForSequenceClassification
2
3tokenizer = AutoTokenizer.from_pretrained("umarkhan0206/medical-intent-classifier")
4model = AutoModelForSequenceClassification.from_pretrained("umarkhan0206/medical-intent-classifier")