Multi-modal data, deep learning, clustering, predictive modeling, type 2 diabetes, dementia, clustering
Date
2025
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
Wearable sensors are increasingly utilized in healthcare to collect physiological and clinical data, enabling various tasks to enhance diagnosis, prognosis, and intervention strategies. Deep-learning models play a crucial role in processing such data due to their ability to extract complex features. However, many existing approaches focus on single-modality data, making them insufficient for scenarios where static and time-series data coexist. This dissertation presents novel deep-learning solutions for multi-modal healthcare data, with a focus on wearable sensors, addressing predictive and clustering problems. ☐ In the predictive tasks, I developed a neural network that integrates static demographic and lab data with time-series measurements from wearable sensors to forecast the progression of type 2 diabetes. Building upon this, I designed an advanced deep-learning model that leverages both convolutional and recurrent layers to capture intra- and inter-sensor dependencies, enabling a more accurate understanding of dynamic patterns in time-series data. This model significantly improves the prediction of diabetes-related outcomes by extracting and combining patterns from diverse sensor modalities. ☐ In clustering tasks, I proposed a supervised method to classify motor behavior into clinically meaningful clusters for predicting acute health events such as falls and delirium in older adults. I further advanced this work by introducing an unsupervised clustering framework tailored for multi-modal data. This approach jointly optimizes objectives for static and dynamic components, enabling the identification of high-risk individuals based on mobility and cognitive patterns, with potential applications in long-term care settings. ☐ Posterior collapse is a key challenge in representation-based time-series clustering, where latent variables become uninformative during training, causing the KL divergence to vanish and the model to ignore the latent space. This issue undermines clustering performance by preventing the model from learning meaningful data representations. To address this, I proposed an information-aware recovery mechanism that predicts collapse by monitoring mutual information and KL divergence, pauses training, and restores the model to a stable state with adjusted distributions. By semi-randomly reallocating data points, the model avoids retracing the same training path, improving the reliability and performance of clustering algorithms for unsupervised learning. ☐ This dissertation advances the application of deep learning in healthcare by integrating static and dynamic data for prediction, improving clustering methodologies, and addressing critical challenges in representation learning. These works aim for a foundation of more reliable and effective AI-driven healthcare analytics.
Description
Keywords
Deep learning, Health data, Wearable sensors, Diabetes, Healthcare analytics