Validation of Propofol PK/PD Models Against Bispectral Index Recordings During Anaesthesia and Exploratory Analysis of EEG Entropy Measures.

In total intravenous anaesthesia, target-controlled infusion systems rely on pharmacokinetic/pharmacodynamic (PK/PD) models to predict the effect of propofol, a hypnotic drug commonly used in this procedure, and guide anaesthetic delivery. Among the available models, those proposed by Marsh et al., Schnider et al., and Eleveld et al. are the most widely used in clinical practice, although their predictive performance may vary substantially across patient populations and clinical conditions. In parallel, electroencephalography (EEG)-based monitoring methods such as the Bispectral Index are routinely employed to assess the hypnotic component of anaesthesia, despite several known limitations related to inter-individual variability and signal interpretation. Entropy-based EEG measures have attracted increasing interest as alternative descriptors of anaesthetic brain dynamics. The aim of this work was to evaluate the predictive performance of the Marsh, Schnider, and Eleveld propofol PK/PD models through Bispectral Index-based validation in two clinical datasets, and to perform an exploratory comparison between the Bispectral Index and EEG-derived entropy indices. For Marsh and Schnider, Bispectral Index predictions were obtained using the pharmacodynamic mapping proposed by Martin-Mateos et al., while the original Eleveld PK/PD formulation was used directly. Predicted Bispectral Index values were compared with recorded Bispectral Index signals in a private dataset derived from Linassi et al., including 59 patients, and in the public VitalDB database, including 331 patients. Model performance was evaluated using per-patient bias, mean absolute error, root mean square error, and standard deviation of the prediction error. Additional analyses investigated the effects of age, obesity, and surgical context on prediction performance. Finally, Permutation Entropy and Entropy of Difference were computed from EEG recordings and compared with Bispectral Index dynamics across anaesthesia stages. Across both datasets, Eleveld showed the best overall predictive performance, with lower median prediction errors than Marsh and Schnider. In the Linassi dataset, Eleveld also showed the bias closest to zero, although a tendency towards more negative bias values was observed in VitalDB. Age-related differences were particularly evident for Eleveld, with increased prediction error and more negative bias observed in elderly patients. By contrast, obesity appeared to negatively affect Marsh and Schnider more markedly, while Eleveld maintained lower prediction error in obese patients within the VitalDB dataset. Exploratory analyses further suggested that variability in prediction error was associated not only with demographic factors but also with clinical context and intraoperative conditions. Entropy-based EEG indices followed the expected transitions across anaesthesia stages and showed smoother and more regular trajectories than the Bispectral Index. In agreement with previous literature, Permutation Entropy also appeared relatively stable across age groups compared with conventional EEG-derived indices. Permutation Entropy and Entropy of Difference exhibited a very strong correlation at low embedding dimension, suggesting that both indices capture closely related properties of EEG dynamics during anaesthesia. Overall, Eleveld was the best-performing PK/PD model evaluated, while persistent limitations related to age and inter-individual variability remained. In addition, the observed behaviour of Permutation Entropy and Entropy of Difference suggests that entropy-based EEG measures may provide useful complementary information to conventional Bispectral Index monitoring and may represent promising candidates for future pharmacodynamic modelling of anaesthetic brain states.