BIOINFORMATICS SEMINAR SERIES
https://bioinformatics.udel.edu/seminar
Prediction of Behavioral Determinants of Health Using Domain-Specific BERT Embeddings and BiLSTM: A Study on the MIMIC-III Dataset
Saad Althabiti
Ph.D. student, Bioinformatics Data Science
University of Delaware
Abstract: This study aimed to enhance the prediction of Behavioral Determinants of Health (BDoH) from medical records by integrating domain-specific BERT embeddings with BiLSTM, while also addressing the challenge of class imbalance in the dataset. We explored methods like oversampling, undersampling, and adjusting class weights to mitigate bias towards the majority class, which can impair model performance. Our proposed approach, utilizing Bio-ClinicalBERT embeddings with BiLSTM, was compared against other state-of-the-art models and demonstrated superior performance, achieving the highest F1 scores across five BDoH labels. The study highlights the effectiveness of this combined approach in improving classification accuracy and robustness, offering a framework that could be applied to other text classification tasks and advancing the field of health informatics.
Bio: Saad Althabiti is a Ph.D. student in Bioinformatics Data Science. His academic journey includes the attainment of a bachelor’s degree in Health Informatics from Saudi Arabia and a master’s degree in Health Informatics, with a specialization in Data Science Track, from the University of Pittsburgh in 2020. His current research concentrated on text mining within electronic health records, intending to discover diverse clinical outcomes. He is currently co-advised by Dr. Shanker and Dr. Wu.
