University of North Texas
Modeling Electronic Health Records: Interpretable Sequential Approaches for Enhanced Multimodal Embeddings and Future Clinical Outcome Prediction with Time-Aware Irregular Intervals Handling
Abstract
dc:descriptionWith the rapid adoption of Electronic Health Records (EHR), vast amounts of longitudinal and multimodal healthcare data are now available, offering significant opportunities for predictive modeling and clinical decision support. EHR data includes structured elements (e.g., lab results, vital signs) and unstructured clinical notes, capturing a comprehensive patient profile. However, challenges such as irregular time intervals, missing values, and heterogeneous modalities necessitate specialized deep learning models to harness their full potential. This dissertation presents innovative deep learning architectures that enhance the representation, interpretability, and predictive accuracy of multimodal EHR data. To predict Alzheimer’s disease (AD) progression, we developed PPAD, a framework integrating longitudinal and cross-sectional patient data. PPAD employs recurrent neural networks (RNN) and incorporates age at each visit to address time irregularities, significantly improving long-term prediction accuracy. Building on this, we developed TA-RNN, an attention-based time-aware RNN that includes a time embedding layer and a dual-level attention mechanism to enhance interpretability. TA-RNN demonstrated strong performance in predicting AD progression and in-hospital mortality. We further developed CAAT-EHR, a Transformer-based generative embedding model that leverages self- and cross-attention to create task-agnostic longitudinal embeddings. An autoregressive decoder ensures temporal consistency by predicting future visits. Models trained on CAAT-EHR embeddings outperformed those using raw or baseline-processed EHR data across multiple clinical tasks. Collectively, PPAD, TA-RNN, and CAAT-HER, highlight the value of time-aware, interpretable, and generalizable deep learning approaches for improving clinical outcome prediction using multimodal EHR data.
Degree
thesis:*- Grantor dc:publisher
- University of North Texas
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Al Olaimat, Mohammad Ahmad
- Contributors dc:contributor
-
- Serdar Bozdag
- JungHwan Oh
- Mark Albert
- Rajeev Azad
Subjects
dc:subject × 4Rights
dc:rights- Statement dc:rights
-
- Public
- Al Olaimat, Mohammad Ahmad
- Copyright
- Copyright is held by the author, unless otherwise noted. All rights Reserved.
- Language dc:language
- English
Identifiers
dc:identifier.*- Identifier
-
local-cont-no: submission_4340
https://digital.library.unt.edu/ark:/67531/metadc2443349/
ark: ark:/67531/metadc2443349