Continual Multi-Disease Detection in Smart Healthcare Using Wearable Medical Sensors

Abstract

Physical illnesses and mental health problems can not only affect individual well-being but also significantly impact the operation of global society. To prevent a pandemic, detecting diseases promptly is one of the most critical tasks in the healthcare domain. However, disease detection has been a stressful and time-consuming process for patients and the general public for centuries. Traditionally, patients have to visit healthcare providers, undergo a series of physical examinations, and then wait for days to receive their diagnostic results. This paradigm is inadequate for providing efficienthealth monitoring for personal well-being. Fortunately, modern advances in machine learning (ML) and wearable medical sensors (WMSs) offer unprecedented solutions for enabling real-time disease detection. WMSs can collect a wealth of physiological data and vital signs from disease-specific patients. Developers can then develop and train a sophisticated ML model on such data to enable efficient and accurate real-time disease detection. However, conventional ML-driven disease-detection methods rely on customizing individual models for each disease and its associated WMS data. Such methods cannot adapt to new data domains or new classification classes quickly. 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. In this thesis, we address the challenges mentioned above by introducing continual learning (CL) to smart healthcare in three directions:• Continual learning in all scenarios: We enable a compact ML model to continually learn new knowledge in the WMS data domain for smart healthcare applications, where training data become available sequentially from non-stationary distributions, by introducing a replay-style CL algorithm. For example, a trained model needs to learn new knowledge from WMS data collected in a different country (domain-incremental CL), from a different classification class of patients (class-incremental CL), or from a completely different disease (task-incremental CL). • Past-agnostic continual learning: While most existing CL algorithms require access to learned knowledge, either in the form of preserved data or distilled information, to achieve satisfactory performance, we propose a past-agnostic CL framework that enables an ML model to continually acquire new knowledge without the aid of stored data or information. Therefore, the proposed framework is applicable to off-the-shelf ML models that utilize WMS data for smart healthcare applications. • Continual fine-tuning with foundation models: We present a novel method to develop Transformer-based foundation models for smart healthcare applications in the WMS data domain. With the well-pre-trained foundation model, we enable continual fine-tuning for downstream disease-detection tasks with various parameter-efficient fine-tuning (PEFT) algorithms. First, we introduce DOCTOR, a multi-disease detection CL framework based on WMSs. DOCTOR enables domain-, class-, and task-incremental CL for multi-disease detection with a multi-headed deep neural network (DNN) and a replay-style CL algorithm. It leverages a data preservation (DP) algorithm and a synthetic data generation (SDG) module to perform replay-based CL. The multi-headed DNN enables DOCTOR to detect multiple diseases simultaneously based on user WMS data. In complex CL scenarios, DOCTOR achieves 1.43× better average test accuracy, 1.25× better F1-score, and 0.41 higher backward transfer than the naive fine-tuning framework, with a small model size of less than 350KB. Next, we present PAGE, a domain-incremental adaptation strategy with past-agnostic generative replay for smart healthcare. PAGE performs generative replay CL without the need for preserved data or information from prior domains. It utilizes new WMS data and the current model to generate synthetic data that distills the learned knowledge embedded in the current model weights. In addition, we introduce interpretability in PAGE by incorporating an extended inductive conformal prediction (EICP) method to provide statistical guarantees for disease-detection results. PAGE achieves highly competitive performance against state-of-the-art (SOTA) along with superior scalability, data privacy, and feasibility. Furthermore, PAGE enables a nearly 75% reduction in clinical workload with the help of EICP. 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 on a large dataset of WMSdata exclusively collected from healthy individuals. We pre-train the foundation model with a masked data modeling (MDM) objective. We then fine-tune the model using various PEFT methods to adapt it to distinct downstream disease-detection tasks based on WMS data. In addition, COMFORT continually stores the low-rank decomposition matrices obtained from the PEFT algorithms to construct a library for multi-disease detection. Hence, the COMFORT library enables scalable and memory-efficient disease detection on edge devices. COMFORT achieves highly competitive performance while reducing memory overhead by over 52% relative to conventional methods. The frameworks proposed in this thesis lay the groundwork for flexible, scalable, and sustainable medical ML models that enable efficient and effective real-time disease detection in smart healthcare applications.