Modern advances in machine learning (ML) and wearable medical sensors (WMSs) in edge devices have enabled ML-driven disease detection for smart healthcare. Conventional ML-driven methods for disease detection rely on customizing individual models for each disease and its corresponding WMS data. However, such methods lack adaptability to distribution shifts and new task classification classes. In addition, they need to be rearchitected and retrained from scratch for each new disease. Moreover, installing multiple ML models in an edge device consumes excessive memory, drains the battery faster, and complicates the detection process.

To address these challenges, we first introduce DOCTOR, a multi-disease detection continual learning (CL) framework based on WMSs. It employs a replay-style CL algorithm and a multi-headed deep neural network (DNN). The CL algorithm enables the DNN to continually learn new missions in all CL scenarios where different data distributions, classification classes, and disease-detection tasks are introduced sequentially. The multi-headed DNN enables DOCTOR to detect multiple diseases simultaneously based on user WMS data. To further enable CL in off-the-shelf models, we introduce PAGE, a domain-incremental adaptation strategy with past-agnostic generative replay. PAGE enables generative replay without the aid of any preserved data or information from prior domains. In addition, PAGE introduces interpretability with the extended inductive conformal prediction (EICP) to provide statistical guarantees for detection results and reduce clinical workload. Finally, we present COMFORT, a continual fine-tuning framework for foundation models targeted at consumer healthcare. COMFORT introduces a novel approach for pre-training a Transformer-based foundation model in the WMS data domain. We then fine-tune the foundation model using various parameter-efficient fine-tuning (PEFT) methods, such as Low-Rank Adaptation (LoRA) and its variants, to adapt it to various down­ stream disease-detection tasks that rely on WMS data. COMFORT continually stores the low-rank decomposition matrices obtained from the PEFT algorithms to construct a library for multi-disease detection. The COMFORT library enables scalable and memory-efficient disease detection on edge devices. Our CL frameworks pave the way for personalized and proactive solutions to efficient and effective early-stage disease detection for smart healthcare applications.

Adviser: Niraj Jha

Zoom Mtg: https://princeton.zoom.us/j/94210252495?pwd=seFgBr1bfEQiWEluit1obakZgiI6F4.1

Meeting ID: 942 1025 2495
Passcode: 339312