COVID-19 From Symptoms to Prediction: A Statistical and Machine Learning Approach

Abstract

Abstract During the COVID-19 pandemic, analysis of patients’ data played a vital role in developing precautions, medications, and vaccination strategies. In this regard, data reported by hospitals and medical institutes is considered one of the reliable sources for any investigation. The use of recent technologies such as machine learning provides a platform to transform such data into meaningful insight that can help decision-makers to prepare future strategies. This study mainly focused on developing prediction models to predict the age group of COVID-19 patients using different attributes by applying statistical and Machine Learning (ML) approaches. The study was conducted in two different phases. Firstly, statistical tests such as ANOVA and t-test were applied to investigate relationships between different variables. Secondly, multiple ML models were applied to predict patients’ age groups based on symptom data. For this, Decision Tree, Naïve Bayes, KNN, Gradient Boosted Trees, and Random Forest models were trained for prediction. In addition, to enrich the performance of the prediction model bagging, boosting, and stacking ensemble approaches are used. The statistical results clearly suggested a significant association among five common symptoms in datasets. Moreover, the results of ML implementation indicated that ensemble approaches such as boosting, bagging, and stacking can enhance the prediction accuracy significantly. Overall, the gradient boosting trees (GBT) with bagging marginally outperformed (0.6628), but in some cases boosting and stacking proved strong techniques and can generate better ensembles than bagging. For example, K-Nearest Neighbors (KNN) and Naïve Bayes (NB) without ensemble recorded (0.529) and (0.554) accuracy, but with the stacking model the performance of both was enhanced significantly and accuracy measured at 0.63 and 0.622 respectively. This study shows that ML ensemble approaches can enrich the performance of the prediction model. The results of this study can be useful for medical authorities to prepare and implement strategies and precaution guidelines for different age groups and recorded symptoms.