Predictive modelling of clinical outcomes in acute tonsillitis based on microbiome analysis and machine learning algorithms

Acute tonsillitis is a common disease with high clinical variability. Traditional approaches based on clinical scores (e.g., Centor) are often insufficient for accurately predicting individual outcomes. The aim of the study was to determine the significance of integrating clinical parameters and oropharyngeal microbial composition data to construct a predictive model for disease duration and symptom severity using the random forest method. Fifty-two patients with acute tonsillitis were examined. Bacteriological analysis of oropharyngeal swabs, clinical assessment using the Centor score, and rapid testing for streptococcal and viral infections were performed. Random forest and linear discriminant analysis models were constructed and compared. The random forest model demonstrated higher accuracy compared to linear discriminant analysis, especially for predicting pain intensity (overall accuracy 81.8% vs 55.0%). For disease duration, the accuracy of the random forest was 72.7% vs 75.0% for linear discriminant analysis. Feature importance analysis revealed that integrating microbiome indices (pathogen/commensal ratio – Pathogen_ratio) with the Centor clinical score significantly improved predictive ability. Disease duration was associated with bacterial aetiology (positive streptococcal test) and smoking status, while pain intensity correlated with microbial dysbiosis parameters. The combination of clinical and microbiological data in machine learning models improves the accuracy of disease progression prediction and can be used to develop personalised treatment approaches