Efficient, robust and uncertainty aware mobile health

Abstract

Sensor-equipped smartphones and wearables are transforming various mobile applications, including health monitoring. As the difference between consumer health wearables and medical devices begins to soften, it is now common for a single wearable device to monitor a range of medical risk factors. Potentially, these devices could give patients direct access to personal analytics that can contribute to their health, facilitate preventive care, and aid in the management of ongoing illness. This data can be fed into machine learning algorithms aiming to help practitioners better assess the progress of the patient or other aspects of the disease evolution. Therefore, it is essential to have accurate model predictions and, equally significantly, a better understanding of what happens within the machine learning model suggesting a prediction. Deep learning is considered a key element in analyzing these new data types. However, traditional deep learning models cannot capture the predictive uncertainty leading to overconfident predictions, and, ultimately, they are less robust in real-world applications. Bayesian deep learning or other non-Bayesian probabilistic techniques can naturally quantify such uncertainty. Still, they do come with quite a few drawbacks, the biggest one being the fact that they tend to be computationally heavy. This dissertation attempts to address the challenges above by developing new efficient-by-design frameworks for robust and uncertainty aware mobile health. Firstly, we introduce a framework that enables already trained deep learning models to generate uncertainty estimates on edge computing platforms with no need for fine-tuning or re-training strategies. This simple and system-efficient solution is built to provide predictive uncertainty for convolutional neural networks based only on one forward pass and a negligible number of additional matrix multiplications. Next, we evaluate our approach on multiple mobile sensing datasets and show that it can provide reliable uncertainty estimates and accurate predictions with very little computation and memory overhead. Next, we provide a new interpretation of early exit neural networks as an implicit ensemble of weight-sharing sub-networks from which predictive uncertainty can be estimated. Our approach can be applied to any feed-forward deep learning architecture. Empirical evaluation of several medical imaging and biosignal datasets and state-of-the-art architectures demonstrates strong performance in accuracy and uncertainty metrics as well as computation gain, highlighting the benefit of combining multiple structurally diverse models that can be jointly trained. Finally, acknowledging that in real-world deployments, deep learning-based mobile health applications may encounter malign inputs that can have a catastrophic impact, we propose an attack-agnostic adversarial mitigation approach based on the previously described early exit ensemble framework. We show that this approach can guarantee robustness generalizable to various white box and universal adversarial inference time attacks, increasing the accuracy of vulnerable targeted deep learning models while providing mitigation comparable to adversarial training but without the computational burden. The frameworks devised in this thesis suggest promising future directions for efficient uncertainty estimation in mobile health research. Furthermore, designing robust deep learning models has increasingly proved to be extremely relevant to health-critical mobile applications. By addressing several challenges in this problem space, these contributions take a few steps toward building machine learning systems that can be safely deployed in real-life settings.