DOI: 10.20935/acaddrug8393 ISSN: 3071-2521

Pharmacometric generative stochastic modeling of patient-reported outcome measures

Kuteesa R. Bisaso, Karungi S. Bisaso, Karyaburo R. Kadada, Jackson K. Mukonzo, Ene I. Ette
Show PDF Cite
Introduction: Patient-reported outcome measures (PROMs) capture the patient’s own perspective on their health, illness, and therapeutic effects on the illness. PROM data are inherently high-dimensional, discrete, and interdependent, challenging standard models that rely on restrictive assumptions and often collapse item-level information. This study, therefore, investigates the use of restricted Boltzmann machines (RBM) to jointly model, simulate, and interpret multidimensional PROM data within a pharmacometric framework.

Materials and methods: A mixed-variate RBM was applied to longitudinal neuropsychological impairment data from 157 HIV-positive patients receiving efavirenz. The model jointly represented binary symptom items, an ordinal mini-mental state examination (MMSE) measure, efavirenz mid-dose concentrations, and clinical variables (CD4 cell count and viral load) using an energy-based formulation. Parameters were estimated via persistent contrastive divergence. Model performance was evaluated using reconstruction error, pseudolikelihood, free-energy stability, and predictive accuracy on a held-out dataset, alongside simulation-based diagnostics including visual predictive checks. The model was used to derive a variable importance ranking for all the PROM items, clinical variables, and drug concentrations.

Results: The RBM captured joint dependencies across PROM items and covariates. On the held-out test set, the model demonstrated good conditional predictive performance, with log-loss ranging from 0.01 to 0.72, Brier scores from 0.00 to 0.26, and mean squared error (MSE) from 0.01 to 0.50. Excellent calibration was achieved at week 12 (t84), particularly for sleepwalking, tactile, and visual hallucinations. Visual predictive checks showed good agreement between observed and simulated symptom trajectories. Variable importance analysis indicated that mid-dose concentrations were not more predictive of post-baseline PROMs than clinical variables and baseline PROMs. Therapeutic drug monitoring simulations revealed limited impact of concentration capping on neuropsychological outcomes.

Conclusions: Generative RBMs provide a flexible and minimally assumptive framework for pharmacometric PROM analysis, enabling joint modeling and simulation of complex symptom data. This approach complements traditional methods and supports the use of baseline clinical state over single exposure metrics for prediction.

More from our Archive