Symptom‐Based Machine Learning Framework for Early Syndrome Coronavirus 2 ( SARS ‐ CoV ‐2) Infection Prediction

Early and accurate identification of SARS‐CoV‐2 infection remains a critical challenge, especially in limited‐infrastructure healthcare settings where access to molecular diagnostics is restricted. This study presents a symptom‐based machine learning framework for rapid prediction of COVID‐19 infection using routinely collected clinical indicators. The dataset used in this work is openly accessible at https://redivis.com/datasets/rxta‐4v35cgyz , enabling full transparency and reproducibility. After comprehensive preprocessing, recursive feature elimination with a random forest estimator identified nine key predictors, including SpO 2 , fever, loss of smell/taste, shortness of breath, dry cough, fatigue, age, diabetes status, and respiratory rate. Multiple supervised learning algorithms—logistic regression, SVM, k‐NN, decision tree, random forest, and a stacked ensemble—were trained and evaluated using 10‐fold cross‐validation. The stacked ensemble model outperformed all baseline classifiers, achieving an accuracy of 0.87, balanced sensitivity (0.88) and specificity (0.86), and the lowest error rate (0.12), demonstrating strong robustness and generalization by leveraging complementary strengths of the base learners. The proposed framework is computationally lightweight, delivers prediction times under 50 ms, and is suitable for deployment in real‐time clinical decision‐support systems. These results highlight its potential as a scalable, rapid, and cost‐effective screening tool for early COVID‐19 detection in resource‐constrained environments. Furthermore, SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model‐agnostic Explanations) analyses confirm that the model's predictions are grounded in clinically valid symptom patterns, with SpO 2 and fever identified as the dominant global predictors, enhancing clinical transparency and trust. Statistical significance testing via McNemar's test confirmed that the stacking ensemble's performance improvements over all baseline classifiers are statistically significant ( p < 0.05 for all pairwise comparisons).