Evaluating time influence over performance of machine-learning-based diagnosis: a case study of Covid-19 Pandemic in Brazil

MARQUES, Julliana Gonçalves, GUEDES, Luiz Affonso and DA COSTA ABREU, Marjory (2022). Evaluating time influence over performance of machine-learning-based diagnosis: a case study of Covid-19 Pandemic in Brazil. International Journal of Environmental Research and Public Health, 20 (1): 136.

Preview

PDF Da Costa-Abreu-EvaluatingTimeInflucence(VoR).pdf - Published Version Creative Commons Attribution. Download (871kB) | Preview

Abstract

Efficiently recognising severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) symptoms enables a quick and accurate diagnosis to be made, and helps in mitigating the spread of the coronavirus disease 2019. However, the emergence of new variants has caused constant changes in the symptoms associate with COVID-19. These constant changes directly impact the performance of machine-learning-based diagnose. In this context, considering the impact of these changes in symptoms over time is necessary for accurate diagnoses. Thus, in this study, we propose a machine-learning-based approach for diagnosing COVID-19 that considers the importance of time in model predictions. Our approach analyses the performance of XGBoost using two different time-based strategies for model training: month-to-month and accumulated strategies. The model was evaluated using known metrics: accuracy, precision, and recall. Furthermore, to explain the impact of feature changes on model prediction, feature importance was measured using the SHAP technique, an XAI technique. We obtained very interesting results: considering time when creating a COVID-19 diagnostic prediction model is advantageous.

Item Type:	Article
Uncontrolled Keywords:	Toxicology
Identification Number:	https://doi.org/10.3390/ijerph20010136
SWORD Depositor:	Symplectic Elements
Depositing User:	Symplectic Elements
Date Deposited:	22 Dec 2022 10:25
Last Modified:	22 Dec 2022 10:25
URI:	https://shura.shu.ac.uk/id/eprint/31197

Actions (login required)

View Item

Downloads

Publication Metrics