LEADD: Learning Efficient and Accurate Disease Diagnoses

Abstract

Many people who end up suffering from a serious medical condition do not have easy access to or never proactively seek out professional treatment. This problem is especially pronounced for mental health: mental illness will affect one in four people at some point in their lives, yet two-thirds of those never seek out the healthcare resources that could help to alleviate their suffering. More generally, in today's economy, healthcare resources such as physicians and hospital infrastructure are coming under increasingly heavy strains, pointing towards the need for a new model of healthcare delivery that can handle increased demand, for mental health and nearly every other disease. The recent proliferation of smartphones and Internet-of-Things (IoT) devices can address this challenge. Smartphones and other IoT devices enable a continuous and passive mechanism to collect and analyze physiological signals, which are precursors for real-time disease diagnosis and health monitoring. Diagnostic models deployed on these devices have the potential to detect all manners of diseases at an early stage. In this thesis, we explore a framework for mental disorder diagnosis and monitoring using efficient neural networks. We also propose a new method for synthesizing efficient neural networks which can be incorporated into any disease diagnosis framework to boost its efficacy and robustness. Through experiments with self-collected datasets and those downloaded from the web, we demonstrate the high accuracy that can be achieved by neural networks when operating directly on patients' physiological signals, and also the benefits in terms of memory and computation reduction when an emphasis on efficiency is adopted during training. We also show that our novel network synthesis method is competitive with the state-of-the-art while improving upon model compactness and efficiency. These results verify the feasibility of neural networks for real-time disease monitoring and treatment, and point towards further research on developing these robust systems for eventual deployment in real-world scenarios.