Building an extensible, AI-augmented ecological momentary assessment platform

In the rapidly evolving landscape of data-driven research, mobile ecological momentary assessment (EMA) has emerged as a powerful tool for real-time data collection. TigerAware, an advanced system designed for the user-friendly creation and deployment of surveys on mobile devices, has undergone significant enhancements to become a multi-faceted platform. This thesis presents several innovations that leverage advancements in large language models (LLMs) and foundation models to augment TigerAware's capabilities and advance the field of mobile EMA. We developed novel methods for automating the qualitative analysis of unstructured text responses collected through EMA. Our LLM-driven approach, which includes a Keypoints Extractor module and a Keypoints Relevance Evaluator, demonstrated superior performance in extracting concise, relevant keywords and accurately linking free text arguments to their pertinent points. Additionally, we introduced Argument2Code (A2C), a method that utilizes LLMs to automate inductive coding in qualitative data analysis by generating comprehensive codebooks. These advancements significantly enhance the automation and sophistication of the coding process in qualitative research. Furthermore, we proposed FENAP (Foundation Models for Comprehensive Analysis and Prediction of EMA-derived Narratives), which applies the capabilities of foundation models to interpret and analyze EMA data. By transforming raw EMA inputs into structured narratives, FENAP captures detailed behavioral patterns and dynamics, enabling a deeper understanding of human behavior. We implemented this methodology using the diverse UT1000 Project dataset and demonstrated that FENAP outperforms traditional and advanced machine learning baselines in predicting behavioral outcomes. To address critical challenges in privacy protection within EMA, we introduced a novel dataset and a hybrid AI approach that combines DistilBERT and LLaMA models for detecting privacy-sensitive information (PSI). This approach surpassed existing baselines in PSI detection accuracy and provided clearer explanations for model decisions, advancing privacy protection methods in EMA. In addition to these contributions, we made significant enhancements to the user experience, workflow optimization, and software engineering aspects of TigerAware. These improvements include interface refinements, streamlined workflow tools, enhanced communication features, and the incorporation of strong software engineering standards, ensuring TigerAware's reliability, extensibility, and user-friendliness for cutting-edge research. Our contributions underscore TigerAware's evolution into a dynamic and comprehensive EMA platform, ideal for the demands of various research domains. The integration of modern AI technologies, such as LLMs and foundation models, into real-world data collection and analysis platforms demonstrates the immense potential for advancing research in the field of mobile EMA.