Based a machine learning approach to investigate the factors influencing nirmatrelvir/ritonavir exposure in human plasma: a multicenter, observational study.

Objectives: Nirmatrelvir/ritonavir (N/R) is an effective antiviral for treating COVID-19. However, evidence supporting therapeutic drug monitoring (TDM) for N/R remains limited, potentially increasing the risk of adverse reactions and compromising efficacy. This study aims to identify factors influencing N/R plasma exposure and to develop and internally validate a machine learning model for predicting N/R concentrations, thereby supporting individualized therapy. Methods: We retrospectively analyzed data from 139 patients who received N/R at two centers. Baseline clinical and laboratory variables were collected, and steady-state trough concentrations of nirmatrelvir and ritonavir were measured on day 3 of treatment. Logistic regression was used to examine the association between drug concentration and prognosis. After excluding highly correlated features, a random forest model identified key factors affecting drug exposure. An XGBoost regression model was then constructed with the selected features, and its predictive performance was evaluated using mean absolute error (MAE), mean squared error (MSE), root mean squared error (RMSE), and R². Five-fold cross-validation was applied for internal validation. Results: Nirmatrelvir trough concentration was not predictive of patient outcomes (AUC = 0.467). Six factors were consistently identified as important determinants of N/R exposure: estimated glomerular filtration rate (eGFR), creatine kinase (CK), aspartate aminotransferase (AST), alanine aminotransferase (ALT), lymphocyte count (Lymph), and procalcitonin (PCT). Ultimately, the evaluation of the predictive model resulted in a mean absolute error (MAE) of 0.717, mean squared error (MSE) of 1.328, root mean squared error (RMSE) of 1.152, and coefficient of determination (R-squared) of 0.779. The prediction model performs well and can provide risk prediction for medication management for N/R, as well as assist in personalized medication. Conclusions: We identified a set of variables that affect the treatment of N/R through therapeutic drug monitoring and established a machine learning model capable of predicting N/R concentrations with satisfactory performance. These findings provide a basis for integrating TDM with multivariable prediction tools to personalize N/R dosing and improve medication safety.