Using Supervised Machine Learning Techniques To Classify Respiratory Tract Infections

Abstract

OBJECTIVES: Acute respiratory tract infections (RTIs) are a leading cause of morbidity and mortality among children under five years of age, and the fourth leading cause of mortality in adults globally. Misclassification of RTIs remains a significant challenge due to inconsistent diagnostic practices, particularly in low- and middle-income countries (LMICs), especially within sub-Saharan Africa. This study aimed to develop a clinical decision support system using a machine learning approach to assist healthcare providers in the accurate diagnosis and classification of RTIs. METHODS: This was a secondary data analysis research study of data collected from a retrospective observational cohort study of respiratory tract infections (RTI) and non-RTI cases, namely acute febrile disease of unknown cause (AFDUC) and gastrointestinal tract infection (GTI), surveyed between July 2018 and December 2020. The primary objectives of the study were to develop, fine-tune, and evaluate a predictive model capable of classifying RTI cases using training datasets that combined data from RTI and AFDUC cases, as well as RTI and GTI cases. The machine learning workflow involved two key phases: data pre-processing and feature engineering, both of which were essential in preparing the data for model development, testing, validation, and final evaluation aimed at achieving accurate diagnostic classification. Classification and prediction models were built using three machine learning algorithms: artificial neural networks (ANN), random forest (RF), and support vector machines (SVM). K-fold cross-validation was employed to optimize and fine-tune the models for improved performance. Model evaluation was conducted using performance metrics derived from confusion matrices, which facilitated the assessment of the receiver operating characteristic (ROC) curves and area under the curve (AUC). Furthermore, feature importance analysis was carried out to identify the most influential variables contributing to robust model performance, while variable correlation analysis was used to explore interactions among the independent variables. RESULTS: Amongst the syndromes, RTI had highest overall numbers of cases enrolled, largely from age group 1 year and below with 42.0% of cases when compared to AFDUC and GTI. For correlation measurements, few variables correlated positively with value between 0.55 and 0.71, particularly age group which correlated with chest pain, headache and HIV status in all models. The evaluation of each model and classifier reached performance metric scoring of above 0.8 for accuracy, precision, specificity and F1-score, which was desirable. The classified RTI ROC curve shape was robust with the expected shape, which was able to measure the AUC above 0.92 for three classes and AUC above 0.95 for two classes. CONCLUSIONS: In this study, we differentiated RTIs from other infections that may present with similar signs and symptoms—such as cough, fever, and elevated temperature—observed during patient admission to the hospital. The predictive models developed and presented in this study provide accurate estimations for distinguishing RTI cases from AFDUC and GTI cases. These findings offer important evidence that RTIs can be effectively and efficiently screened using machine learning approaches, without the limitations typically associated with time, manpower, and cost constraints.