Authors:
A. Sandhya, G. Deena, Mohamed Ahmed Hassan Youssef, Shivani Arora, Yahya Ali Saleh Al-Quhali, Ara Kiarash, Cao Yu

Addresses:
Department of Computer Science and Engineering in Artificial Intelligence and Machine Learning, SRM Institute of Science and Technology, Ramapuram, Chennai, Tamil Nadu, India. College of Medicine, Zagazig University, Zagazig, Sharqia Governorate, Egypt. Department of Data and Analytics, Infosys Limited, Edmonton, Alberta, Canada. Department of Basic Sciences, Amran Community College, Amran, Yemen. Department of Electrical and Chemical Engineering, McMaster University, Hamilton, Ontario, Canada. Department of Computing and Information Systems, Singapore Management University, 81 Victoria Street, Singapore.

Abstract:

The clinical narratives contained in EHRs are long, unstructured, with dense information content, and difficult to classify due to uneven length, structure, terminology, and information density. This work presents a hybrid machine learning system that uses classifiers to classify medical transcription, focusing on data efficiency, interpretability, and robustness. The pipeline combines natural language preprocessing, Domain-informed feature extraction, Hashing Vectorizer representation and BERT representation for context. Five features identified in the original study document length, bigram type, relationship words, word count, and noun/verb counts are included to capture complementary lexical, syntactic, and domain information. The Medical Transcriptions Dataset consists of 80% training and 20% testing data. The reported evaluation includes 2321 test samples. Test-set accuracy is about 74.2%, weighted precision is about 74.4%, weighted recall is about 74.2%, and weighted F1 score is about 72.7%. The average macro F1 is around 50.0%, which shows significant class-to-class variation. Many low-support categories have low or zero F1 values, while several high-support categories have strong F1 values. After 6 epochs, the training loss drops from 0.67 to 0.46, and the reported validation F1 rises from 66% to 77% and ultimately levels off. These results show that a hybrid representation of heterogeneous medical text is feasible, but also indicate significant class imbalance and limited minority-class examples. The updated framework thus suggests balanced learning, external validation, systematic ablation, robustness testing, and clinically assessed interpretability before real-world deployment.

Keywords: Electronic Health Records; Clinical NLP; Hashing Vectorizer; Medical Text Classification; Machine Learning; Clinical Text Mining; Medical Text Classification.

Received on: 24/08/2025, Revised on: 07/11/2025, Accepted on: 26/11/2025, Published on: 03/06/2026

DOI: 10.69888/FTSBE.2026.000793

FMDB Transactions on Sustainable Biomedical Engineering, 2026 Vol. 1 No. 2, Pages: 130-142

• Views : 24
• Downloads : 6

Download PDF