A hybrid machine learning/pharmacokinetic approach improves predictive performance in model-informed precision dosing
Jasmine H. Hughes, Ron J. Keizer
Objectives: Model-informed precision dosing (MIPD) typically leverages empirical Bayes estimates of pharmacokinetic (PK) parameters, allowing for individualization of treatment at the patient bedside. These algorithms combine information about the patient, in the form of therapeutic drug monitoring samples or collection of pharmacodynamic biomarkers, with prior knowledge of drug PK, in the form of a population PK model. The model prior acts as an anchor, preventing PK parameter estimates from deviating too far from the population averages. Often, this smoothing effect provides desirable protection against noisy clinical data, until sufficient evidence accumulates to justify extreme PK estimates. However, some patients may not be well described by a model prior, and if no alternative models are available, little guidance exists in how to individualize dosing in these situations. We investigated the use of “flattened priors” (FP), in which the weight of the model priors is decreased during maximum a posteriori (MAP) Bayesian estimation. We further trained machine learning (ML) models to predict when FP would outperform conventional MAP for a particular clinical decision point.
Methods: De-identified data entered or imported into the InsightRX Nova platform over the course of routine clinical care of adult patients treated with vancomycin formed the basis of this analysis. Patients were included if 2+ drug serum levels were collected over the course of treatment (N = 4679 patients, 9052 drug levels). For each sample, patient PK parameters were estimated using MAP Bayes estimation with either the interindividual variability magnitude (IIV) as reported (MAP) or by inflating IIV of all parameters eightfold (FP), which effectively reduces the penalty associated with eta estimates during maximum likelihood estimation. These parameters were then used to predict the value of the subsequent TDM level. This iterative prediction process was repeated for three previously published population PK models [1-3]. Prediction error was assessed with root mean squared error (RMSE). Patient records were randomly split into training, cross validation, and testing data sets. ML models (XGBoost, logistic regression, penalized logistic regression) were developed to predict when FP should be applied based on 29 features thought to be predictive of the use of FP.
Results: FP outperformed MAP in 42-55% of parameter estimations, with a theoretical reduction in RMSE of 19-23% if the best of MAP or FP had been used. Of the ML models tested, the XGBoost models performed the best (accuracy: 75-77%, specificity: 81-86%), reducing RMSE by 12-22% relative to MAP. Predictive features showed good agreement between the models; the most important features were the cumulative bias in prediction error and the most recent prediction residual. A minimal logistic regression model trained on just these two features approximated the performance of the more complex models (accuracy: 65%, specificity: 76%), reducing RMSE by 5-18% relative to MAP. Minimal hybrid models trained on pairwise combinations of the three PK models and evaluated on the third “unseen” PK model also performed quite well (accuracy: 59-66%, specificity: 71-81%, reduction in RMSE: 5-15%).
Conclusion: FP improves predictive performance for a subset of patients, and these instances can be predicted by ML models through hybrid ML/PK algorithms. The consensus in predictive features and the generalizability of these ML models to unseen PK models suggests this approach may also apply to other drugs and patient populations. Applications of ML to pharmacometrics have become increasingly popular in recent years however ML models trained to predict PK exposure parameters, such as drug levels or area under the curve, typically preclude simulation of patient response to alternative drug regimens or mechanistic insight into patient PK. This hybrid ML/PK approach uniquely leverages the ability of ML to learn from many patients and extract information about an individual from a rich feature set, while maintaining the mechanistic insight and interpretability of PK models. This approach enables clinicians to dose patients not well described by available PK models.
 Thomson, A. H., Staatz, C. E., Tobin, C. M., Gall, M. & Lovering, A. M. Development and evaluation of vancomycin dosage guidelines designed to achieve new target concentrations. J. Antimicrob. Chemother. 63, 1050–1057 (2009).
 Goti, V., Chaturvedula, A., Fossler, M. J., Mok, S. & Jacob, J. T. Hospitalized Patients with and Without Hemodialysis Have Markedly Different Vancomycin Pharmacokinetics: A Population Pharmacokinetic Model-Based Analysis. Ther. Drug Monit. (2018).doi:10.1097/FTD.0000000000000490
 Buelga, D.S., Mar Fernandez De Gatta, M. Del, Herrera, E. V., Dominguez-Gil, A. & García, M. J. Population pharmacokinetic analysis of vancomycin in patients with hematological malignancies. Antimicrob. Agents Chemother. (2005).doi:10.1128/AAC.49.12.4934-4941.2005
 McComb, M., Bies, R. & Ramanathan, M. Machine Learning in Pharmacometrics: Opportunities and Challenges. Br. J. Clin. Pharmacol. (2021).doi:10.1111/bcp.14801