UC Irvine

Artificial intelligence is transforming healthcare from episodic assessment into continuous and personalized monitoring. In practice, a monitoring system rarely ends at the sensor: physiological measurements, contextual observations, model updates, or learned representations may pass from a wearable device to a phone, an edge gateway, a cloud service, or an institutional platform for storage, collaborative learning, clinical interpretation, and model reuse. These flows make intelligent monitoring scalable and adaptive, but they also expand the privacy surface beyond conventional identifiers. This dissertation develops a continuous research program around this operational tension: it begins with user-aware data acquisition, adds privacy-preserving adaptive learning, and then progressively examines what sensitive information can remain in the data and artifacts that move through the healthcare AI lifecycle. The first part of the dissertation focuses on adaptive health monitoring. A context-aware active reinforcement-learning framework learns when to request ecological momentary assessments by considering prediction uncertainty together with time, response history, and smartphone context. The framework progresses from offline label selection to real-time deployment and personalized stress detection, showing how a monitoring system can collect informative labels while reducing unnecessary user burden. Once these data are collected, limited on-device samples and the need to improve models across users motivate collaborative learning infrastructure. A differential private federated transfer-learning framework therefore combines public-data pretraining, decentralized user adaptation, and protected aggregation. This approach supports learning from sparse and distributed health data while reducing the need to centralize raw observations and limiting disclosure through shared model updates. Federated learning and protected aggregation reduce exposure at a specific stage, but they do not eliminate the broader need to share, process, and reuse healthcare information. Raw records may later be de-identified for research, representations may be transferred to downstream applications, and pretrained encoders may be deployed through cloud or institutional interfaces. The second part therefore follows the residual privacy risk as information changes form along this pipeline. An interpretable analysis of electrocardiogram (ECG) data demonstrates that physiological morphology itself can reveal demographic attributes and participant identity across heterogeneous datasets. TransECG extends this analysis to transformer representations and uses model-intrinsic attention to localize identity-related evidence within physiologically meaningful ECG regions. The dissertation then moves from measuring biometric information to evaluating its operational consequences: a linkage attack shows how de-identified ECG records can be connected across data sources under partial adversarial knowledge, while a membership-inference audit shows that reusable ECG foundation encoders can reveal whether an individual or cohort participated in pretraining through exposed scores or embeddings. Together, these studies demonstrate that removing explicit identifiers or decentralizing training does not make the remaining lifecycle artifacts inherently anonymous: privacy leakage can arise from the signal itself, from learned representations, from cross-dataset sharing, and from deployed model interfaces. The central conclusion is that secure healthcare intelligence must be adaptive at two levels: it must adapt interaction and learning to the user, and it must adapt privacy safeguards to the changing attack surface across the data and model lifecycle. No single mechanism, such as de-identification or decentralized training, is sufficient on its own. Trustworthy healthcare AI instead requires burden-aware sensing, privacy-preserving learning, interpretable risk analysis, realistic adversarial evaluation, and continuous auditing of the interfaces through which data and models are shared.