Infectious Disease Symptom-Based Multiclass Prediction Using Bernoulli Naïve Bayes

Abstract

The symptoms of respiratory infectious diseases like influenza, common cold and COVID-19 are similar, which makes it difficult to differentiate them in the initial screening. In this paper, a multiclass infectious disease prediction system is proposed based on a Bernoulli Naïve Bayes classifier for the three disease classes, based on structured binary symptom inputs. The study is based on the adapted CRISP-DM methodology which includes data understanding, data preprocessing, modelling and evaluation processes. Six publicly available datasets from Kaggle, Mendeley Data, and GitHub were harmonized into a unified ten-feature symptom schema and reduced through duplicate-pattern limiting to 9,170 records. The classifier was developed from scratch in Python, and tested with stratified 5-fold cross validation, with SMOTE only applied to the training folds to avoid synthetic samples from contaminating the test fold. The mean accuracy of the model was 84.79% across the five folds, while the weighted precision, recall, and F1-score were 0.8763, 0.8479, and 0.8551, respectively. The results of the confusion matrix analysis showed that the highest recall was for influenza, whereas the highest precision was for common cold and the lowest precision was for COVID-19, which is due to the similarity of the symptoms between the three classes. The results show that a lightweight and interpretable probabilistic classifier, along with harmonized binary symptom information, can be used effectively as a decision support tool for early-stage multiclass infectious disease screening.