TY - JOUR
T1 - Detecting the most critical clinical variables of COVID-19 breakthrough infection in vaccinated persons using machine learning
AU - Daramola, Olawande
AU - Kavu, Tatenda Duncan
AU - Kotze, Maritha J.
AU - Kamati, Oiva
AU - Emjedi, Zaakiyah
AU - Kabaso, Boniface
AU - Moser, Thomas
AU - Stroetmann, Karl
AU - Fwemba, Isaac
AU - Daramola, Fisayo
AU - Nyirenda, Martha
AU - van Rensburg, Susan J.
AU - Nyasulu, Peter S.
AU - Marnewick, Jeanine L.
N1 - Publisher Copyright:
© The Author(s) 2023.
PY - 2023/12
Y1 - 2023/12
N2 - Background: COVID-19 vaccines offer different levels of immune protection but do not provide 100% protection. Vaccinated persons with pre-existing comorbidities may be at an increased risk of SARS-CoV-2 breakthrough infection or reinfection. The aim of this study is to identify the critical variables associated with a higher probability of SARS-CoV-2 breakthrough infection using machine learning. Methods: A dataset comprising symptoms and feedback from 257 persons, of whom 203 were vaccinated and 54 unvaccinated, was used for the investigation. Three machine learning algorithms – Deep Multilayer Perceptron (Deep MLP), XGBoost, and Logistic Regression – were trained with the original (imbalanced) dataset and the balanced dataset created by using the Random Oversampling Technique (ROT), and the Synthetic Minority Oversampling Technique (SMOTE). We compared the performance of the classification algorithms when the features highly correlated with breakthrough infection were used and when all features in the dataset were used. Result: The results show that when highly correlated features were considered as predictors, with Random Oversampling to address data imbalance, the XGBoost classifier has the best performance (F1 = 0.96; accuracy = 0.96; AUC = 0.98; G-Mean = 0.98; MCC = 0.88). The Deep MLP had the second best performance (F1 = 0.94; accuracy = 0.94; AUC = 0.92; G-Mean = 0.70; MCC = 0.42), while Logistic Regression had less accurate performance (F1 = 0.89; accuracy = 0.88; AUC = 0.89; G-Mean = 0.89; MCC = 0.68). We also used Shapley Additive Explanations (SHAP) to investigate the interpretability of the models. We found that body temperature, total cholesterol, glucose level, blood pressure, waist circumference, body weight, body mass index (BMI), haemoglobin level, and physical activity per week are the most critical variables indicating a higher risk of breakthrough infection. Conclusion: These results, evident from our unique data source derived from apparently healthy volunteers with cardiovascular risk factors, follow the expected pattern of positive or negative correlations previously reported in the literature. This information strengthens the body of knowledge currently applied in public health guidelines and may also be used by medical practitioners in the future to reduce the risk of SARS-CoV-2 breakthrough infection.
AB - Background: COVID-19 vaccines offer different levels of immune protection but do not provide 100% protection. Vaccinated persons with pre-existing comorbidities may be at an increased risk of SARS-CoV-2 breakthrough infection or reinfection. The aim of this study is to identify the critical variables associated with a higher probability of SARS-CoV-2 breakthrough infection using machine learning. Methods: A dataset comprising symptoms and feedback from 257 persons, of whom 203 were vaccinated and 54 unvaccinated, was used for the investigation. Three machine learning algorithms – Deep Multilayer Perceptron (Deep MLP), XGBoost, and Logistic Regression – were trained with the original (imbalanced) dataset and the balanced dataset created by using the Random Oversampling Technique (ROT), and the Synthetic Minority Oversampling Technique (SMOTE). We compared the performance of the classification algorithms when the features highly correlated with breakthrough infection were used and when all features in the dataset were used. Result: The results show that when highly correlated features were considered as predictors, with Random Oversampling to address data imbalance, the XGBoost classifier has the best performance (F1 = 0.96; accuracy = 0.96; AUC = 0.98; G-Mean = 0.98; MCC = 0.88). The Deep MLP had the second best performance (F1 = 0.94; accuracy = 0.94; AUC = 0.92; G-Mean = 0.70; MCC = 0.42), while Logistic Regression had less accurate performance (F1 = 0.89; accuracy = 0.88; AUC = 0.89; G-Mean = 0.89; MCC = 0.68). We also used Shapley Additive Explanations (SHAP) to investigate the interpretability of the models. We found that body temperature, total cholesterol, glucose level, blood pressure, waist circumference, body weight, body mass index (BMI), haemoglobin level, and physical activity per week are the most critical variables indicating a higher risk of breakthrough infection. Conclusion: These results, evident from our unique data source derived from apparently healthy volunteers with cardiovascular risk factors, follow the expected pattern of positive or negative correlations previously reported in the literature. This information strengthens the body of knowledge currently applied in public health guidelines and may also be used by medical practitioners in the future to reduce the risk of SARS-CoV-2 breakthrough infection.
KW - COVID-19
KW - Explainable AI
KW - J&J vaccine
KW - Machine learning
KW - Pfizer vaccine
KW - XGBoost
KW - breakthrough infection
KW - deep multilayer perceptron
KW - logistic regression
KW - vaccination
UR - http://www.scopus.com/inward/record.url?scp=85176403960&partnerID=8YFLogxK
U2 - 10.1177/20552076231207593
DO - 10.1177/20552076231207593
M3 - Article
C2 - 37936960
AN - SCOPUS:85176403960
SN - 2055-2076
VL - 9
SP - 1
EP - 23
JO - Digital Health
JF - Digital Health
ER -