| dc.description.abstract | The utility of machine learning for enhancing human well-being and health has risen to the core discussion in both research and real-world application in today’s technological front-line. The fastgrowing artificial intelligence industry has innovated health-related applications. Inversely the realworld challenges in accurate implementation of mobile and clinical health solutions have necessitated advancements in theoretical and algorithmic development. The increasing rate of digitizing medical records in hospitals has enable artificial intelligence with abundant data to train. In return the trained models have shown to lower the uncertainty in clinical decision making. However, as always, with opportunities comes new challenges. We have observed many discrepancies of compatibility
between sophisticated machine learning models and the nuanced clinical needs, such as fairness, personalized treatment through precise phenotyping and data shift in longitudinal medical records. In modeling complex and heterogeneous health record-based machine learning and then extrapolating through remote health applications, I have identified the need for advanced, multimodal models that continually learn risk representation from varied, heterogeneous data sources. Broadly, I recognize a few gaps in different levels currently blocking us from enhancing continual machine learning for health: 1) domain-, 2) class-, and 3) personal-level heterogeneity in realworld healthcare data.
With the three proposed aims, I will target at using machine learning in a more robust and generalizable way towards real-world biomedical data and to enhance not only the prediction accuracy but also the interpretation of the results. The proposed work will largely benefit both machine learning domain as well as human-centered computing application. This dissertation will mostly introduce and elaborate how to bridge the gap between algorithm in the lab and the open-world challenges and hopeful will spur more research onto this interdisciplinary problem and bring about real world improvements. | |
