Development and validation of an artificial intelligence-based model for predicting teicoplanin plasma concentrations
Qinghua Zhang1, Qi Zhang2, Banglong Wang1
1The First Affiliated Hospital of Dalian Medical University, 222 Zhongshan Road, DalianLiaoning, 116021, China.
Introduction:
Teicoplanin is widely used to treat pulmonary infections, particularly in critically ill patients with gram-positive bacterial infections. However, its plasma concentration varies substantially between individuals owing to heterogeneity in renal function, comorbidities, and concomitant medications, leading to subtherapeutic or toxic exposures. Therapeutic drug monitoring (TDM) remains the standard approach for individualized dosing; however, its application in intensive care unit (ICU) is limited by resource and timing constraints. Hence, developing reliable predictive models to estimate teicoplanin plasma concentrations could enhance the precision and efficiency of TDM, and support pharmacist-led dosing decisions.
Aim:
This study aimed to develop and validate machine learning (ML)-based models to predict the teicoplanin plasma concentration in ICU patients with pulmonary infections using real-world clinical data, thereby advancing personalized antibiotic therapy in clinical pharmacy practice.
Method:
This retrospective cohort study was conducted between June 2018 and September 2023 in ICU patients receiving teicoplanin therapy at a tertiary hospital in China. After the univariate analysis to exclude irrelevant factors, sequential forward selection was performed to identify the optimal feature subset. The dataset was divided into a training set (80%) and a test set (20%), and ten ML algorithms were developed to predict teicoplanin plasma concentrations. Model performance was evaluated by ten-fold cross-validation of the training set and validated using an independent external cohort.
Results:
A total of 270 eligible patients were included in the training and test sets, and additional 118 patients formed the external validation cohort. Seven variables were identified as the most predictive features: daily dose, diabetes, hemodialysis, imipenem, human serum albumin, urea, and red blood cell count. Among the ten algorithms tested, the TabNet model achieved the best predictive performance on the test set (R2 = 0.88, RMSE = 3.24, MAE = 2.64, MAPE = 17.88%, ± 30% accuracy = 81.54%) and maintained robust external validation (R2 = 0.79, ± 30% accuracy = 85.59%).
Conclusion:
This study developed a TabNet model based on real-world data, which can accurately and interpretably predict the teicoplanin plasma concentration of patients with pulmonary infection in the ICU. The derived online prediction tool provides references for the application of artificial intelligence-assisted precision medicine in antibacterial treatment and the optimization of drug monitoring for ICU patients.
Insights
Machine learning accurately predicts teicoplanin levels in ICU patients with pulmonary infections. This AI tool aids personalized antibiotic therapy and optimizes drug monitoring for better patient outcomes.
Area of Science:
- Pharmacology
- Artificial Intelligence
- Intensive Care Medicine
Background:
- Teicoplanin is crucial for treating Gram-positive bacterial pulmonary infections in critically ill patients.
- Individual teicoplanin plasma concentrations vary widely, risking subtherapeutic or toxic levels.
- Current therapeutic drug monitoring (TDM) in ICUs faces resource and timing limitations.
Purpose of the Study:
- To develop and validate machine learning (ML) models for predicting teicoplanin plasma concentrations in ICU patients.
- To leverage real-world clinical data for personalized antibiotic therapy.
- To support clinical pharmacy practice in optimizing teicoplanin dosing.
Main Methods:
- Retrospective cohort study of ICU patients receiving teicoplanin (June 2018 - September 2023).
- Feature selection identified key predictors: daily dose, diabetes, hemodialysis, imipenem, albumin, urea, and RBC count.
- Ten ML algorithms were trained and validated, with TabNet showing superior performance.
Main Results:
- The TabNet model achieved high predictive accuracy (R²=0.88, 81.54% ±30% accuracy) on the test set.
- External validation confirmed robust performance (R²=0.79, 85.59% ±30% accuracy).
- Key predictors included patient factors and concomitant treatments influencing teicoplanin levels.
Conclusions:
- A validated TabNet model accurately predicts teicoplanin plasma concentrations in ICU patients.
- This AI-driven tool facilitates precision medicine for antibacterial treatment.
- The model supports optimized drug monitoring and pharmacist-led dosing decisions in critical care.
