Why did individuals seeking COVID-19 advice return to internet hospitals for follow-up in the early pandemic: Insights from machine learning

Abstract

BACKGROUND: Exploring the factors influencing individuals seeking coronavirus disease 2019 (COVID-19) consultation to choose internet hospitals for follow-up care after their initial in-person visit is crucial for optimizing medical resource allocation and enhancing future infectious disease control. This study, anchored in the core constructs of the Health Belief Model, aims to systematically examine the key determinants influencing patients' decision to pursue a second COVID-19 consultation via an Internet hospital following an initial face-to-face visit. METHODS: This study included 1,055 individuals seeking advice about COVID-19 who consulted an internet hospital during the early outbreak. Model training and evaluation were conducted using stratified five-fold cross-validation; in each fold, 844 subjects were allocated to the training set and 211 to the test set. We employed a full-feature baseline alongside four feature selection methods-analysis of variance (ANOVA), Boruta, least absolute shrinkage and selection operator (Lasso), and all subsets regression (ASR), and three resampling techniques: Oversampling, Oversampling & Undersampling, and Artificial Synthesis Dataset. Eight machine learning models-logistic regression (LR), support vector machine (SVM), k-nearest neighbors (KNN), random forest (RF), fully connected neural network (FCNN), and extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), and categorical boosting (CatBoost)-were constructed to compare performance metrics and rank variable importance in the best-performing model. RESULTS: The impact of different feature selection methods and resampling techniques on model performance varied. Statistically significant differences were observed in the performance of each model in terms of the area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPRC) (AUROC: P < 0.001, AUPRC: P < 0.001). Following Boruta, the RF model using the Artificial Synthesis Dataset demonstrated the highest AUROC at 0.946 ± 0.027, whereas the LightGBM model trained on the original imbalanced dataset achieved the highest AUPRC at 0.864 ± 0.070. The key factors influencing the decision to use internet hospitals for follow-up consultations included taking medication on their own or as prescribed by offline doctors before the second visit, chronic respiratory diseases history, contact, consultation purposes, fatigue, cough, and sore throat. CONCLUSIONS: Machine learning may help identify factors influencing the choice of internet hospitals for follow-up consultations. Integrating the Health Belief Model enables a more nuanced understanding of individuals' online consultation behaviors. Optimizing triage services could enhance the quality and accessibility of internet hospitals, offering valuable insights for future public health events.

Ähnliche Arbeiten