Semester of Graduation
Spring 2026
Degree Type
Dissertation
Degree Name
Computer Science
Department
Computer Science
Committee Chair/First Advisor
Md Abdullah Al Hafiz Khan
Abstract
Behavioral health plays a pivotal role in individuals’ overall quality of life. Timely identification of behavioral health concerns is essential for building a supportive community. Real-world data sources such as police reports, clinical notes, and social media contain valuable cues about behavioral health concerns. However, manual screening is time-consuming, and the shortage of qualified professionals further delays identification. Although deep learning and natural language processing techniques have shown great potential for analyzing textual data, these data sources are unstructured and context-dependent, making it challenging for traditional models to identify cases accurately. Moreover, support from subject matter experts (SMEs) is essential. However, existing human-in-the-loop systems primarily utilize SMEs for data labeling while overlooking their contextual knowledge.
To address this gap, this dissertation developed human-AI collaboration methodologies that are accurate, scalable, and adaptive to human insights. This dissertation progressed through several phases, each addressing core requirements for effective collaboration. In the initial phase, we developed a domain-enhanced attention network that integrates SME-provided domain knowledge with self-attention mechanisms to improve performance on noisy real-world data. In addition, we developed a framework using prompt-based learning and graph attention networks to capture relationships among behavioral health evidence. In the next phase, we focused on interpretability by designing framework that utilizes multiple level of features along with multi-level attention for effective identification. In the final phases we developed collaborative and adaptive human-AI collaboration frameworks that integrates expert feedback and achieves 82.23% accuracy. Extensive evaluations on real-world datasets demonstrated the efficacy and effectiveness of our proposed approaches.