UCLA

Over the past decades the dramatic increase in computational resources coupled with the advent of machine learning and artificial intelligence has profoundly impacted sensor technology. Together, these trends present a new opportunity for data-driven computational sensor design, where acquisition hardware is fundamentally changed to “lock-in” to the optimal sensing data with respect to a user defined cost function. This dissertation addresses this emerging opportunity, referred to herein as computational sensing, with a focus on applications in biophotonics and diagnostics where there is currently high-demand for cost-effective and

compact sensing systems for democratizing global health. Therefore, several examples of computational sensing systems will be presented, all of which leverage machine learning and data-driven design principles to enable low-cost, compact, and sparse implementations well-suited for the point-of-care among other settings.

Firstly, I will discuss a multiplexed paper-based vertical flow assay and reader for biomarker quantification at the point-of-care. I will show how deep learning and statistical analysis of the multiplexed sensor response can be leveraged to improve accuracy, dynamic range, and limit-of-detection, greatly enhancing the sensing capabilities beyond that of the traditionally employed lateral flow assay. Additionally, I will demonstrate feature selection techniques which can serve as a toolbox for iterative assay development, sensor design, and cost-performance optimization. Two applications in point-of-care diagnostics will be presented with this platform, one being a clinical study for early-stage Lyme Disease diagnoses, and the other, a clinical study for cardio-vascular risk assessment by the high-sensitivity C-Reactive Protein test. Secondly, Localized plasmon resonance (LSPR) sensors and their corresponding readers will be discussed as computational sensing systems for label-free biomolecular sensing. Through machine learning of spectral characteristics, I will present how the corresponding LSPR sensor reader can be jointly designed with the LSPR sensor to optimize read-out accuracy while minimizing cost by computationally selecting from cost effective illumination optics. Lastly, a fundamentally different hardware design for spectroscopy, enabled by deep learning, will be demonstrated and discussed for application-specific performance advantages in terms of acquisition rate, form factor, and cost.