Composite Importance Sampling (CIS) for continuous learning in precision dosing
Ron J Keizer (1), Dominic MH Tong (1), Jasmine H Hughes (1), Srijib Goswami (1)
(1) InsightRX, San Francisco, CA
Introduction: Continuous learning (CL) in model-informed precision dosing (MIPD) can improve accuracy in forecasting and dosing advice.[1,2,3] CL leverages data collected during routine use of clinical decision support (CDS) tools to optimize predictions in future patients. It has shown promise especially in settings where the appropriateness of the model is unknown or suspected to be poor, and especially improves empirical (a priori) dosing.[3] Production-grade non-supervised automation is challenging for gradient- and sampling-based NLME methods due to potentially intensive computational loads at scale, the inability to incrementally update model parameter estimates, potential estimation instabilities requiring manual model adjustments, the inability to dynamically assess covariate effects, and the inability to perform federated model estimation. Here, we introduce Composite Importance Sampling (“CIS”), an adaptation of the importance sampling algorithm, that has the potential to resolve these challenges for continuous learning approaches in MIPD.
Objectives: Evaluate the ability of CIS to learn from routinely collected patient data and improve predictive accuracy of MIPD platforms.
Methods: The CIS algorithm consists of three distinct steps: 1) sampling of model parameter sets from a proposal distribution, performed only once; 2) calculation of the individual likelihood for all parameter sets in an incoming single-patient dataset, performed each time new data becomes available; 3) aggregation of the individual likelihoods into cumulative likelihoods and calculation of the posterior densities using importance weighting [4], performed whenever a posterior estimate is desired. Updated parameter estimates can then be obtained from the weighted posterior density and used in the CDS tool. The CIS algorithm was implemented in R based on the open source PKPDsim framework.[5]
Simulated PK scenarios (n=1 to 100 patients) were used to evaluate whether misspecification of a model could be detected by CIS on realistic TDM data. A hypothetical drug with linear 1-compartment PK was used, in which various biases were introduced: a) missing a categorical covariate effect (genotype), and b) missing a continuous covariate (allometric scaling) effect, and c) biased initial parameter estimates.
A separate simulated scenario evaluated whether CIS could be applied to a neutropenia model used in an MIPD setting.[2,6] Bias was introduced in PD parameters mean transit time (MTT, +20%) and drug effect (slope, +50%).
Lastly, a retrospective real data analysis (n=323 patients) was used to evaluate whether a literature model for gentamicin in neonates [7] previously shown to be biased could be rendered more predictive using CIS in a test data set (n=138 patients).
For all scenarios, posterior parameter estimates were obtained cumulatively on growing datasets from 1 to 100 patients (simulated data) and 1 to 323 patients (retrospective data). Accuracy and bias in prediction of future drug concentrations were assessed using RMSE and MPE after each patient was added to the dataset.
Results: In the simulated PK scenario, the missing covariate effects as well as the biased estimates for the PK parameters were clearly and accurately identified by CIS. In the simulated PK-PD scenario, although still biased, posterior estimates for both PD parameters were closer to the true value than the prior value.
In the retrospective scenario based on routine gentamicin TDM data, CIS identified and quantified bias in the literature model parameters, and reduced RMSE by >30% relative to the initial model. Retrospective analysis showed that CIS would have been able to improve accuracy early: even after inclusion of data from 10 patients RMSE was reduced by ~20% relative to the initial RMSE.
Conclusion: CIS allows for stable, unsupervised, iterative and federated analysis of incoming data to approximate the posterior distribution and re-estimate population-level model parameters in MIPD. Furthermore, CIS allows ad hoc exploration of covariate effects without the need for model re-estimation. Further exploration is needed to evaluate the appropriateness of CIS for further MIPD scenarios.
References:
- Keizer RJ et al. CPT-PSP 2018. DOI: 10.1002/psp4.12353
- Maier C et al. CPT-PSP 2020. DOI: 10.1002/psp4.12492
- Hughes JH et al. CPT 2021. DOI: 10.1002/cpt.2088
- Dosne AG et al. JPKPD 2016. DOI: 10.1007/s10928-016-9487-8
- https://cran.r-project.org/web/packages/PKPDsim/index.html
- Wallin et al. Basic Clin Pharmacol Toxicol 2010. DOI: 10.1111/j.1742-7843.2009.00520.x
- De Cock R et al. Pharm Res 2014, DOI: 10.1007/s11095-014-1361-z
