Power efficient machine learning-based hardware architectures for biomedical applications

The future of critical health diagnosis will involve intelligent and smart devices that are low-cost, wearable, and lightweight, requiring low-power, energy-efficient hardware platforms. Various machine learning models, such as deep learning architectures, have been employed to design intelligent healthcare systems. However, deploying these sophisticated and intelligent devices in real-time embedded systems with limited hardware resources and power budget is complex due to the requirement of high computational power in achieving a high accuracy rate. As a result, this creates a significant gap between the advancement of computing technology and the associated device technologies for healthcare applications. Power-efficient machine learning-based digital hardware design techniques have been introduced in this work for the realization of a compact design solution while maintaining optimal prediction accuracy. Two hardware design approaches, DeepSAC and SABiNN have been proposed and analyzed in this work. DeepSAC is a shift-accumulator-based technique, whereas SABiNN is a 2's complement-based binarized digital hardware technique. Neural network models, such as feedforward, convolutional neural nets, residual networks, and other popular machine learning and deep neural networks, are selected to benchmark the proposed model architecture. Various deep compression learning techniques, such as pruning, n-bit (n = 8,16) integer quantization, and binarization on hyper-parameters, are also employed. These models significantly reduced the power consumption rate by 5x, size by 13x, and improved the model latency. For efficient use of these models, especially in biomedical applications, a sleep apnea (SA) detection device for adults is developed to detect SA events in real-time. The input to the system consists of two physiological sensor data, such as ECG signal from the chest movement and SpO2 measurement from the pulse oximeter to predict the occurrence of SA episodes. In the training phase, actual patient data is used, and the network model is converted into the proposed hardware models to achieve medically backed accuracy. After achieving acceptable results of 88 percent accuracy, all the parameters are extracted for inference on edge. In the inference phase, reconfigurable hardware validated the extracted parameter for model precision and power consumption rate before being translated onto the silicon. This research implements the final model in CMOS platforms using 130 nm and 180 nm commercial CMOS processes.

Name thesis:degree_name

Ph. D.

Level thesis:degree_level

Doctoral

Discipline thesis:degree_discipline

Electrical and computer engineering (MU)

Grantor dc:publisher

University of Missouri--Columbia

Year dc:date.issued

2023

Author dc:creator

• Hassan, Omiya

Advisor dc:contributor.advisor

• Islam, Syed Kamrul

Language dc:language.iso

eng, English

DOI

https://doi.org/10.32469/10355/96085

OAI identifier oai:identifier

oai:mospace.umsystem.edu:10355/96085