Development of a Mobile COVID and Disease Diagnostic Tool Using Machine and Deep Learning

Abstract

Ever since the onset and subsequent development of the novel COVID-19 virus in December 2020, the main challenge faced by every country across the globe is adequate testing.[1] Without adequate, accessible, and widespread testing, it becomes nearly impossible to trace the virus and ascertain its prevalence in a community. While many relatively “rapid-PCR” tests have developed, the turn around rates from giving the test to receiving the result are still between 48-72 hours in many parts of the world.[2] This time lag between getting tested and getting a result significantly damages any efforts to contain the virus. Furthermore, in many third-world and developing countries, accessibility – particularly financial accessibility – is still an issue.[3] Many people still cannot afford to get a COVID test, especially in parts of the world where these tests are largely conducted by privatized laboratories.[3] Thus, this paper focuses on the development of a novel mobile-based COVID-19 detection/diagnosis tool which serves as a precursor to the PCR test as its primary focus. Doing so will reduce the burden on the testing system as this will “slow” down the flow of people directly lining up to take the PCR test. The idea is for people to use this tool (which can take in both blood-based reports and indicators or purely observable physiological symptoms) to determine an initial probability of COVID. If their risk level is above a certain threshold (which is still yet to be determined experimentally), then they should proceed to take the COVID PCR test. As a secondary goal/extension, this project will develop an “overall” Disease Diagnostic tool overall and display capability to diagnose Diabetes as well as Heart Disease. In the process of doing so, the paper also develops an end-to-end framework for deploying Deep Learning models on Android applications using PyTorch Mobile[4] and Tensorflow Lite[5], which are the mobile-friendly libraries under PyTorch[4] and Tensorflow[6] respectively. Numerous Machine Learning classifiers were trained using Scikit-Learn[7], but due to compatibility issues, MultiLayer Perceptrons[8] were designed in PyTorch and Tensorflow for deployment on the mobile application. The application displays capability of diagnosing COVID using just observable symptoms as well as a full blood work report. It also displays capability of diagnosing Diabetes and Heart Disease.