Mixed effect estimation in deep compartment models: Variational methods outperform first-order approximations

• DOI: 10.1007/s10928-024-09931-w
• Publication Date: 2024-07-04T16:01:58Z
• Authors: Alexander Janssen; Frank C. Bennis; Marjon H. Cnossen; Ron A. A. Mathôt; ; S. H. Reitsma; F. W. G. Leebeek; M. Coppens; K. Fijnvandraat; K. Meijer; S. E. M. Schols; H. C. J. Eikenboom; R. E. G. Schutgens; F. Heubel-Moenen; L. Nieuwenhuizen; P. Ypma; M. H. E. Driessens; I. van Vliet; M. J. H. A. Kruip; S. Polinder; P. Brons; F. J. M. van der ...; K. Fischer; K. van Galen; P. W. Collins; M. Mathias; P. Chowdary; D. Keeling; J. Lock; H. C. A. M. Hazen...; T. Preijers; N. C. B. de Jager; L. Schutte; L. H. Bukkems; M. C. H. J. Goedhart; J. M. Heijdra; L. Romano; W. Al Arashi; M. E. Cloesmeijer; S. F. Koopman; C. Mussert
• Keywords: NONMEM

Abstract This work focusses on extending the deep compartment model (DCM) framework to estimation of mixed-effects. By introducing random effects, predictions can be personalized based drug measurements, enabling testing different treatment schedules an individual basis. The performance classical first-order (FO and FOCE) machine learning variational inference (VI) algorithms were compared in a simulation study. In VI, posterior distributions variables are approximated using whose parameters directly optimized. We found that approximations estimated path derivative gradient estimator version VI highly accurate. Models fit simulated data set FO objective functions gave similar results, with accurate both population covariate effects. Contrastingly, models FOCE depicted erratic behaviour during optimization, resulting parameter estimates inaccurate. Finally, we methods two real-world sets haemophilia A patients who received standard half-life factor VIII concentrates prophylactic perioperative settings. Again, although some presented divergent results. unstable. conclusion, show mixed-effects DCM is feasible. performs conditional estimation, which might lead more results complex method.