Smart Healthcare via Efficient Machine Learning

Abstract

U.S. healthcare spending reached 3 trillion dollars in 2014 and is expected to rise to 5.4 trillion dollars by 2024. Despite this heavy expenditure, preventable medical errors account for around 100,000 deaths/year in U.S. hospitals, only behind heart disease and cancer. Though clinical healthcare has significantly improved general welfare, it remains limited to hospitals/clinics. This will change with the advent of wearable medical sensors (WMSs), which make around-the-clock health monitoring feasible. WMSs collect physiological signals from the body. They can both bend the healthcare costs downwards and reduce medical errors significantly. In this thesis, we take a significant step towards smart healthcare by leveraging the strengths of WMSs and efficient machine learning.

We first propose a novel healthcare framework called the Health Decision Support System (HDSS) that enables disease diagnosis in both in- and out-of-clinic scenarios through the integration of WMS data with clinical decision support systems (CDSSs). HDSS sequentially structures the information framework for daily health monitoring, initial clinical checkup, detailed clinical examination, and post-diagnostic decision support. We have demonstrated the feasibility of HDSS through diagnosis of six diseases aimed at four ICD-10-CM disease categories for which the datasets are publicly available. Just the WMS tier offers diagnostic accuracies ranging from 78% to 99%.

We further evaluate the effectiveness of the WMS based disease diagnosis through DiabDeep – a framework that combines off-the-shelf WMSs and efficient neural networks for pervasive diabetes diagnosis. DiabDeep completely bypasses the computationally-intensive feature extraction stage. It adopts a grow-and-prune paradigm that learns both the architecture and weights of the neural networks for accurate yet efficient inference. Over data collected from 52 participants, DiabDeep yields more than 94% in accuracy when distinguishing among type-1, type-2, and healthy individuals, while reducing the model size (floating-point operations) by up to two orders (one order) of magnitude compared to conventional learning methods.

To effectively monitor 69,000 human diseases on the edge, one has to ensure effi- ciency of inference given very limited communication bandwidth, sensor energy/storage, and training data accessibility upon model deployment. To solve these challenges, we further focus on efficient inference from three different perspectives:

• Communication efficiency. We first study how smart and conventional sensors can work collaboratively along the IoT hierarchy for efficient inference. We propose a novel hierarchical inference model (HIM) based on hierarchical learning and local inferences. HIM takes advantage of inference already performed on smart sensors, while at the same time accommodating conventional sense- and-transmit sensors in the IoT system. We show that for seven applications, the proposed method helps reduce the amount of transmission by up to 60x against conventional baselines, while improving accuracy.

• Model efficiency. We next focus on improving the efficiency of the inference model. We propose a hardware-guided symbiotic training methodology for compact, execution-efficient, yet accurate neural networks. By leveraging the hardware-impacted hysteresis effect, our multi-granular grow-and-prune training strategy enables the symbiosis of model compactness and accuracy with execution efficiency. For two well-known long short-term memory (LSTM) applications, we have achieved 7-31x parameter reduction, 1.7-5.2x latency reduction, while improving accuracy.

• Data efficiency. One fundamental assumption made by existing methods for efficient inference, including the aforementioned methods, relies on access to original training data. While entities might want to share their models, the training datasets are not only large but difficult to store, transfer, and manage. To tackle this, we enable knowledge transfer from a trained convolutional neural network (CNN) to accommodate new design tasks without using the original data. We introduce DeepInversion, a method that converts random noise into high-fidelity class-conditional images given just a pretrained CNN classifier. Further, we introduce Adaptive DeepInversion which utilizes both the teacher and application-dependent student network to improve image diversity. We demonstrate its applicability to three tasks of immense practical importance – (i) data-free network pruning, (ii) data-free knowledge transfer, and (iii) data-free continual learning.