Development of a Clinical Scoring Tool using Machine-Learning prediction of Monospot positivity in suspected Infectious Mononucleosis

Abstract Infectious mononucleosis is clinically difficult to distinguish from bacterial causes of tonsillitis/pharyngitis. First-line investigations include monospot testing, despite 63% sensitivity in certain cohorts. Therefore, national recommendations include repeating an initial negative monospot within five to seven days. A point-of-care clinical scoring tool could improve clinical outcomes, increase diagnostic accuracy and reduce unnecessary testing. We conducted a retrospective cohort study at Luton and Dunstable University Hospital including patients aged 15-24 presenting with sore throat, lymphadenopathy or fever between 01/01/2021-31/01/2024 who underwent monospot testing. Extracted data included demographics, observations, and laboratory results. Patients were randomly split into training (80%) and testing (20%) cohorts. Eleven parameters were used to develop four predictive models; classical multivariate logistic regression, machine-learning logistic regression with LIBLINEAR approximation, machine-learning decision tree classifier, and a simplified clinical risk-stratification model from machine-learning methods. 278 presentations from 264 patients were included. The machine-learning decision tree classifier demonstrated superior performance, achieving 100% sensitivity, 98.0% specificity and 98.2% diagnostic accuracy using only three parameters: lymphocyte count, neutrophil count and alanine aminotransferase. The simplified clinical risk-stratification model demonstrated 83.3% sensitivity, 98.0% specificity and overall 96.4% accuracy. All four models represent potential methods for developing clinical tools to predict monospot positivity. Our risk-stratification model showed significant promise as an easily memorisable, point-of-care clinical scoring tool. Using this, we propose an alternative diagnostic pathway with early counselling and de-escalation of antibiotics in high-risk cases, and reduced testing in low-risk cases; reducing population-level morbidity and epidemiological spread, whilst improving diagnostic accuracy and conserving resources.