A retrospective two-center study conducted in China has developed and externally validated an interpretable machine learning model to estimate teicoplanin plasma concentrations in adult intensive care unit patients undergoing treatment for bloodstream infections. Teicoplanin serves as a first-line lipoglycopeptide antimicrobial for severe Gram-positive bloodstream infections in intensive care settings due to its lower nephrotoxicity relative to vancomycin. However, clinical management is routinely complicated by extreme pharmacokinetic variability among critically ill individuals. Pathophysiological disruptions common to critical illness, including augmented renal clearance, expanded volume of distribution, hypoalbuminemia, and acute multi-organ dysfunction, disrupt normal antimicrobial disposition. Consequently, standard empiric regimens frequently fail to achieve therapeutic plasma thresholds, risking clinical failure, the emergence of antimicrobial resistance, or drug-related toxicity. To address the clinical challenge of individualized exposure assessment, researchers developed an extreme gradient boosting algorithm, known as XGBoost, to predict measured teicoplanin concentrations using readily available demographic, clinical, and laboratory variables.
The investigation enrolled adult patients admitted to intensive care units across two independent tertiary hospital centers who were diagnosed with bloodstream infections, treated with teicoplanin, and monitored via routine therapeutic drug monitoring. The modeling cohort comprised 260 eligible patients treated between January 2024 and December 2025 at the primary center, which was randomly split at an eight-to-two ratio into a training subset of 208 patients and an internal testing subset of 52 patients. To assess cross-center generalizability under real-world conditions, an independent external validation cohort of 91 eligible patients admitted between October 2024 and December 2025 was established at the second center. Cross-checking of patient identification records confirmed that no individuals contributed data to both cohorts. Eligible participants were restricted to adults receiving an unchanged maintenance dose regimen for at least two to three days prior to therapeutic drug monitoring blood sampling, while individuals who died prior to discharge, pregnant or lactating patients, those with incomplete medical records, or those with concentrations below the analytical limit of quantification were excluded.
Data collection extracted routine clinical variables spanning dosing regimens, baseline demographics, comorbid conditions, supportive intensive care therapies, concomitant medications, and laboratory values from electronic medical record systems. Dosing was quantified as the total daily maintenance dose in grams per day in effect immediately prior to blood sampling, excluding initial loading doses. Plasma teicoplanin concentrations were quantified by clinical laboratories at both institutions using routine clinical immunoassays sharing an identical lower limit of quantification of 3.125 micrograms per milliliter. Because exact sampling intervals relative to preceding drug administrations were not systematically logged in electronic charts, investigators designated the target outcome as the measured teicoplanin plasma concentration on the day of sampling rather than a standardized steady-state pre-dose trough concentration. To prepare the dataset, binary variables with class prevalence below ten percent were removed to eliminate extreme class imbalance, missing values were imputed using a random forest iterative algorithm fitted solely on the training data, and continuous variables were normalized using z-score standardization.
Feature engineering utilized a wrapper-based sequential forward selection algorithm using XGBoost with default parameters as the base evaluator, guided by cross-validated coefficient of determination values across two-fold cross-validation. Starting from an empty feature set, candidate variables were iteratively evaluated and permanently retained when demonstrating performance gains. The cross-validated coefficient of determination exhibited a steep initial increase across the first seven variables, plateaued between eight and twenty-five features, and peaked at a subset of ten variables before experiencing a gradual decline associated with overfitting. The final ten-feature subset consisted of total daily dose, patient gender, diabetes, chronic kidney disease, hemodialysis, blood transfusion, mechanical ventilation, alanine aminotransferase, white blood cell count, and red blood cell count. This feature selection reduced the initial thirty-six candidate predictors to ten, thereby increasing the subjects-per-variable ratio from 5.8 to 20.8 within the training cohort of 208 patients.
Investigators comparatively evaluated ten distinct machine learning and deep learning algorithms using the unified ten-feature set, optimizing hyperparameters through ten-fold cross-validation on the training set. The algorithms comprised k-nearest neighbors, adaptive boosting, XGBoost, gradient boosting decision trees, categorical boosting, light gradient boosting machine, multilayer perceptron, support vector machines, a five-layer multilayer perceptron, and Wide and Deep architectures. Across all candidate architectures evaluated during cross-validation, the XGBoost model demonstrated the most favorable point estimates, achieving a mean coefficient of determination of 0.75 with a standard deviation of 0.09. Training was conducted on a single central processing unit without graphics processing unit acceleration, with individual model fitting durations ranging from 1.0 millisecond for k-nearest neighbors to 12.7 seconds for the Wide and Deep network, while incremental peak memory utilization remained under 59 megabytes across all algorithmic pipelines.
Evaluation on the internal testing set of 52 patients confirmed that the XGBoost model delivered superior point estimates among all evaluated algorithms, recording a root mean square error of 3.22 with a 95 percent confidence interval of 2.62 to 3.77, a mean absolute error of 2.52 with a 95 percent confidence interval of 1.98 to 3.10, and a coefficient of determination of 0.73 with a 95 percent confidence interval of 0.56 to 0.81. Furthermore, the model achieved a 30 percent prediction accuracy, defined as the proportion of estimates falling within 30 percent of observed concentrations, of 78.8 percent with a 95 percent confidence interval of 67.3 to 88.5 percent. When tested on the independent external validation cohort of 91 patients, XGBoost sustained consistent predictive performance, yielding a root mean square error of 3.61 with a 95 percent confidence interval of 3.08 to 4.14, a mean absolute error of 2.89 with a 95 percent confidence interval of 2.48 to 3.35, a coefficient of determination of 0.76 with a 95 percent confidence interval of 0.69 to 0.80, and a 30 percent prediction accuracy of 69.2 percent with a 95 percent confidence interval of 59.3 to 78.0 percent.
Paired bootstrap comparisons based on 1,000 resamplings with Holm-Bonferroni correction demonstrated that XGBoost significantly outperformed five of nine comparator algorithms on the internal test set and seven of nine on the external validation cohort. Its performance remained statistically comparable to several top-performing tree-based ensembles, including AdaBoost, CatBoost, and LightGBM internally, and AdaBoost and gradient boosting decision trees externally. Benchmarking comparisons against classical statistical models, including a dose-only linear regression model and a multivariable linear regression model operating on log-transformed concentrations using the identical ten-feature set, confirmed that the XGBoost model yielded superior accuracy across every evaluated performance metric in both cohorts. Sensitivity analyses evaluating all ten algorithms trained on the full set of thirty-six unselected predictors demonstrated consistent relative performance rankings, while complete-case analyses restricted to patients without imputed values confirmed that missing data imputation did not distort the primary experimental conclusions.
To provide model interpretability and assess the clinical plausibility of predictions, researchers applied the TreeExplainer framework of Shapley additive explanations, computing contribution values across the training cohort. The daily maintenance dose emerged as the dominant predictive feature, with model predictions clustering into distinct dose-related tiers, indicating that categorical maintenance dosing accounted for the majority of explained variance. In addition to dose, hemodialysis and mechanical ventilation demonstrated positive associations with predicted teicoplanin plasma concentrations. The authors emphasized that the positive association with hemodialysis reflects severe underlying renal impairment and markedly reduced intrinsic drug clearance rather than dialytic drug accumulation, given that teicoplanin undergoes primary renal elimination and exhibits variable clearance across diverse dialytic modalities. Similarly, mechanical ventilation was interpreted as an indicator of elevated critical illness severity, marked systemic inflammation, and fluid shifts rather than exerting a direct physiological effect on drug disposition. Conversely, four features, namely gender, diabetes, chronic kidney disease, and blood transfusion, exhibited near-zero mean Shapley values across the full cohort.
Exploratory subgroup analyses conducted across both cohorts evaluated model consistency across strata defined by gender, diabetes, chronic kidney disease, hemodialysis, blood transfusion, and mechanical ventilation. Absolute prediction errors, measured by root mean square error and mean absolute error, remained broadly consistent across most clinical strata. In the internal test set, the root mean square error was virtually identical between patients receiving hemodialysis and those not receiving hemodialysis, recording values of 3.21 and 3.20, respectively. However, the corresponding coefficient of determination was notably lower in the hemodialysis group at 0.58 compared to 0.78 in the non-hemodialysis group, an artifact attributable to narrower between-patient concentration variance rather than degraded absolute predictive precision. In the external cohort, mechanical ventilation status was associated with a statistically significant difference, yielding a root mean square error of 3.08 and 30 percent accuracy of 80.95 percent in ventilated patients compared to 4.04 and 63.27 percent in non-ventilated patients, while the transfusion-negative internal subgroup showed an apparent 100 percent accuracy within a sample of only six patients.
The study authors noted several fundamental methodological constraints that govern how the predictive tool should be applied in practice. Because the retrospective design lacked exact hour-level recording of the interval between dosing and sampling, the model estimates exposure at arbitrary routine sampling points rather than standardized steady-state pre-dose troughs. Furthermore, clinical centers administered fixed absolute doses rather than weight-adjusted regimens, and weight data were incomplete, preventing weight-based dosing calculations. Additionally, the cohort excluded patients who died prior to discharge, introducing potential survivor bias that may yield optimistically biased performance metrics by omitting patients with the greatest physiological volatility. Across both validation sets, 21.2 percent and 30.8 percent of predictions deviated beyond the 30 percent accuracy boundary, an error margin that could misclassify target attainment near critical clinical thresholds of 10 and 20 micrograms per milliliter. Consequently, the authors concluded that this XGBoost model represents a probabilistic, dose-anchored screening adjunct intended to complement conventional therapeutic drug monitoring and pharmacokinetic modeling rather than replace direct analytical concentration measurements or clinician judgment.
Subject of Research: Machine learning prediction of teicoplanin plasma concentrations in critically ill patients
Article Title: An interpretable XGBoost model for predicting measured teicoplanin plasma concentrations in ICU patients with bloodstream infection: a two-center real-world validation study
Article References: Zhou, W., Liu, Y., Ma, Y., Sonam, L., Zhang, M., Xiao, Y., Lyu, C., & Wei, L. (2026). An interpretable XGBoost model for predicting measured teicoplanin plasma concentrations in ICU patients with bloodstream infection: a two-center real-world validation study. Frontiers in Pharmacology, 17, Article 1963274. https://doi.org/10.3389/fphar.2026.1963274
Image Credits: AI Generated
DOI: 10.3389/fphar.2026.1963274
Keywords: Teicoplanin, Therapeutic Drug Monitoring, Intensive Care Unit, Bloodstream Infection, XGBoost, Machine Learning, Clinical Pharmacology, Pharmacokinetics, Shapley Additive Explanations, Hemodialysis, Mechanical Ventilation, Sequential Forward Selection, External Validation, Retrospective Study, Precision Dosing, Critically Ill Patients, Serum Creatinine
News Source: Teresa Odom. (November 15, 2026). Machine Learning Model Accurately Predicts Teicoplanin Concentrations in Intensive Care Patients. Scienmag.



