Toward an end-to-end platform for digital health care

: a smart sensor

Abstract

In this work, it was intended to contribute to the development of a Smart Sensor for targeted biomarkers, a much-needed footstep toward healthcare decentralization driven by thought-out the Point of Care (POC) screening capabilities. The bottleneck for this transition lies in the absence of End-to-End (E2E) platforms capable of collecting biomarkers, compiling on dynamic data lakes, and smoothly distributing the information, hardwiring data producers to data consumers, making full use of automation, data flow, and deep learning existing infrastructures. In this context, a sensor for POC needs to combine affordability, and transparency regarding data acquisition and processing, i.e., the Smart Sensor. The respective main components are (i) embedded digital filters for on-the-fly data curation, (ii) modeling, and (iii) deployment. The biomarkers chosen for this work were hemoglobin, glucose, and TMAO. The first two are markers of the person's general health. TMAO is a marker capable of predicting the risk of heart disease. The sensor chosen was an InnoSpectra NIR-M-R2, as it is affordable, portable, and capable of being controlled by IoT devices. However, these sensors present some barriers, such as the need for an experienced user. To combat that, the Smart Sensor has digital filters for data curation. Prediction models for each biomarker were created using the filtered samples and the PLS algorithm. These models were orthogonally corrected against the drying effect of samples. After that, the calibration was transferred between 2 sensors through transfer by orthogonal projection (TOP). Finally, convolutional neural network (CNN) models were applied to automate the model creation process. Digital filters improved RMSEP up to 58%, showing RMSEP for hemoglobin, glucose, and TMAO of 0.575, 2.399, and 1.185 g/l. Comparing CNN to PLS, there are no significant differences between them. However, CNN has several advantages, such as: does not require pre-processing, and the models are already robust to external effects, eliminating the need for an additional DOE. As such, CNN significantly reduces the complexity and time required to create models. Finally, the digital filters and models were embedded in cloud infrastructure, creating an E2E platform that analyzes these biomarkers in POC.

Date of Award	16 Jan 2023
Original language	English
Awarding Institution	Universidade Católica Portuguesa
Supervisor	António Silva Ferreira (Supervisor) & António de Sousa Barros (Co-Supervisor)