Symptom Based Medical Department Determination Using BERT

Within the umbrella of current developments in Natural Language Processing (NLP), this work tackles the crucial problem of effectively classifying patients based on symptoms, thereby assigning patients an appropriate medical speciality and thus improving clinical decision support models. We fine-tuned and explored the efficacy of state-of-the-art BERT-based models, namely DistilBERT and BERT ${ }_{\text {BASE }}$, for classifying patient symptoms into one of the six categories: internal medicine, cardiology, neurology, gastroenterology, dermatology, and infectious diseases. Our experiments observed that BERT ${ }_{\text {BASE }}$ achieved 98% accuracy, while DistilBERT achieved a slightly higher accuracy of 99% on a tweaked dataset of more than 5,000 items.