Predicting prolonged dalbavancin exposure using machine learning: a validated strategy for individualized redosing

TL;DR

This paper presents a machine learning approach to predict if dalbavancin dosing needs adjustment, improving personalized treatment for infections.

Contribution

A validated machine learning model for individualized redosing decisions using minimal clinical data.

Findings

Support vector machine models achieved over 88% accuracy and 90% sensitivity in predicting dalbavancin exposure.

ML predictions outperformed Bayesian estimation in accuracy and sensitivity, especially reducing false negatives.

The model remained reliable through week 8, the clinically relevant exposure period.

Abstract

Dalbavancin is a long-acting lipoglycopeptide increasingly used off-label for complex Gram-positive infections requiring prolonged therapy. Its extended half-life enables simplified regimens, but interindividual pharmacokinetic variability and pathogen MIC heterogeneity complicate dosing. We developed and externally validated machine learning (ML) models to predict whether dalbavancin plasma concentrations remain above predefined pharmacokinetic/pharmacodynamic targets after two standard 1,500 mg doses (day 1/day 8 or day 1/day 15). Predictions were binary (adequate vs subtherapeutic concentration), directly reflecting the clinical decision to readminister a 1,500 mg dose. Models were trained on simulated PK profiles from a published population PK (popPK) model and evaluated in three independent settings: (i) simulated validation data sets from two alternative published popPK models, (ii) a real-world cohort from Limoges University Hospital (n = 31), and (iii) a secondary cohort from Nantes University Hospital (n = 7). Input features included age, body weight, creatinine clearance, MIC, and a single plasma concentration obtained before the second dose. Support vector machine models achieved high accuracy (>88%) and sensitivity (>90%) across testing sets and clinical validation cohorts. In clinical data sets, no false negatives were observed (limited by sample size), with overall accuracy approaching 95%. Compared with maximum a posteriori Bayesian estimation, ML achieved higher accuracy and sensitivity across validation cohorts, particularly by reducing false negatives. Predictions remained reliable through week 8, the clinically relevant exposure window. This ML-based approach enables early individualized redosing decisions using minimal clinical inputs. By complementing Bayesian forecasting and reducing reliance on serial sampling, it represents a pragmatic strategy to support model-informed precision dosing of dalbavancin.

Figures