# Using machine learning to define mepolizumab treatment response at 2 years in patients with chronic rhinosinusitis with nasal polyps

## TL;DR

This study uses machine learning to predict which patients with chronic rhinosinusitis will respond well to mepolizumab treatment after two years.

## Contribution

The study introduces a machine learning approach to identify biomarkers for predicting mepolizumab treatment response in chronic rhinosinusitis with nasal polyps.

## Key findings

- XGBoost was the most accurate model for predicting super-response to mepolizumab with an ROC-AUC of 0.766.
- Blood Eosinophilia and Blood Neutrophilia were significant predictors of treatment response.
- 44.1% of patients were classified as 'super-responders' after 2 years of treatment.

## Abstract

Using machine learning to identify clinical biomarkers for determining optimal response to mepolizumab in chronic rhinosinusitis with nasal polyps.

Single center retrospective observational study with 84 CRSwNP patients treated with mepolizumab. We evaluated 4 machine learning algorithms: Decision Tree, Logistic Regression, K-Nearest Neighbors and Extreme Gradient Boosting. K-Fold cross-validation incorporating hyperparameter optimization in the process was used to ensure robustness and prevent overfitting.

After 6, 12 and 24 months, SNOT-22, VAS overall symptom score, VAS-smell, asthma control test (ACT) and nasal polyp score (NPS) significantly improved (p < 0.001). 44.1% of patients were classified as “super-responders” after 2-year of Mepolizumab treatment based on EPOS/Euforea criteria. XGBoost emerged as the most accurate for predicting super-response to mepolizumab, achieving an ROC- AUC of 0.766. In contrast, Logistic Regression was the least effective for predicting sustained super-response at 24 months, with an ROC-AUC of 0.628. Significant predictors included Blood Neutrophilia and Blood Eosinophilia where higher baseline scores were linked to higher probabilities of super-response at 24 months. Shapley Additive Explanations were employed to identify the most critical baseline features and to visualize their directional impacts on treatment responses.

Machine learning models, particularly XGBoost, can predict real-world super-response to mepolizumab in severe CRSwNP by identifying key predictors like high baseline BEC, high baseline BNC and AERD comorbidity. These insights have the potential to refine CRSwNP treatment strategies and support clinical decision-making, ultimately enhancing patient outcomes by predicting treatment response prior to starting medication