Bayesian population modeling of drug dosing adherence

The value of adherence information during clinical pharmaceutical trials

BACKGROUND/AIMS

The quality of the evidence used to evaluate a drug's safety and efficacy depends, in part, on how well participants adhere to the prescribed drug-taking regime. There are multiple approaches to measure adherence in clinical trials, varying in their cost and accuracy. We demonstrate a method for evaluating the cost-effectiveness of common adherence monitoring methods, considering the costs and data quality for drugs that differ in how forgiving they are of nonadherence.

METHODS

We propose a simulation approach to estimate the value of evidence about adherence, considering both costs of collection and potential errors in interpreting clinical trial results. We demonstrate the approach with a simulated clinical trial of nitrendipine, a common calcium channel blocker. We consider two trial designs, one using pretrial adherence to "enrich" the trial sample and one without an enrichment strategy. We use scenarios combining high and low values of two key properties of a clinical trial: participant adherence and drug forgiveness.

RESULTS

Under the conditions of these simulations, the most cost-effective adherence monitoring approach depends on both trial participant adherence and drug forgiveness. For example, the enrichment strategy is not cost-effective for the base scenario (high forgiveness and high adherence), but is for other scenarios. We also estimate the effects of evaluable patient analysis, a controversial procedure that excludes nonadherent participants from the analyses, after a trial is completed.

CONCLUSIONS

Our proposed approach can guide drug regulators and developers in designing efficient clinical trials and assessing the impact of nonadherence on trial results. It can identify cost-effective adherence-monitoring methods, given available knowledge about the methods, drug, and patients' expected adherence.

Machine learning-guided, big data-enabled, biomarker-based systems pharmacology: modeling the stochasticity of natural history and disease progression

The incidence of systemic and metabolic co-morbidities increases with aging. The purpose was to investigate a novel paradigm for modeling the orchestrated changes in many disease-related biomarkers that occur during aging. A hybrid strategy that integrates machine learning and stochastic modeling was evaluated for modeling the long-term dynamics of biomarker systems. Bayesian networks (BN) were used to identify quantitative systems pharmacology (QSP)-like models for the inter-dependencies for three disease-related datasets of metabolic (MB), metabolic with leptin (MB-L), and cardiovascular (CVB) biomarkers from the NHANES database. Biomarker dynamics were modeled using discrete stochastic vector autoregression (VAR) equations. BN were used to derive the topological order and connectivity of a data driven QSP model structure for inter-dependence of biomarkers across the lifespan. The strength and directionality of the connections in the QSP models were evaluated using bootstrapping. VAR models based on QSP model structures from BN were assessed for modeling biomarker system dynamics. BN-restricted VAR models of order 1 were identified as parsimonious and effective for characterizing biomarker system dynamics in the MB, MB-L and CVB datasets. Simulation of annual and triennial data for each biomarker provided good fits and predictions of the training and test datasets, respectively. The novel strategy harnesses machine learning to construct QSP model structures for inter-dependence of biomarkers. Stochastic modeling with the QSP models was effective for predicting the age-varying dynamics of disease-relevant biomarkers over the lifespan.

Evaluating digital medicine ingestion data from seriously mentally ill patients with a Bayesian Hybrid Model

The objective of this work was to adapt and evaluate the performance of a Bayesian hybrid model to characterize objective temporal medication ingestion parameters from two clinical studies in patients with serious mental illness (SMI) receiving treatment with a digital medicine system. This system provides a signal from an ingested sensor contained in the dosage form to a patient-worn patch and transmits this signal via the patient’s mobile device. A previously developed hybrid Markov-von Mises model was used to obtain maximum-likelihood estimates for medication ingestion behavior parameters for individual patients. The individual parameter estimates were modeled to obtain distribution parameters of priors implemented in a Markov chain-Monte Carlo framework. Clinical and demographic covariates associated with model ingestion parameters were also assessed. We obtained individual estimates of overall observed ingestion percent (median:75.9%, range:18.2–98.3%, IQR:32.9%), rate of excess dosing events (median:0%, range:0–14.3%, IQR:3.0%) and observed ingestion duration. The modeling also provided estimates of the Markov-dependence probabilities of dosing success following a dosing success or failure. The ingestion-timing deviations were modeled with the von Mises distribution. A subset of 17 patients (22.1%) were identified as prompt correctors based on Markov-dependence probability of a dosing failure followed by a dosing success of unity. The prompt corrector sub-group had a better overall digital medicine ingestion parameter profile compared to those who were not prompt correctors. Our results demonstrate the potential utility of a Bayesian Hybrid Markov-von Mises model for characterizing digital medicine ingestion patterns in patients with SMI.

A Receiver Operating Characteristic Framework for Non-adherence Detection Using Drug Concentration Thresholds-Application to Simulated Risperidone Data in Schizophrenic Patients

Non-adherence to antipsychotic medication is a primary factor in disease relapse in schizophrenic patients. We sought to evaluate if plasma concentrations of the antipsychotic risperidone can be used as a predictor of treatment adherence and to identify the optimal plasma concentration threshold to reliably distinguish between adherent and non-adherent patients. A population pharmacokinetic model was used to simulate plasma risperidone steady-state trough concentrations in 1000 virtual patients, where 60% of the patients were 100% adherent to their medication, while 40% of the patients were non-adherent to their medication. The probability of adherence was assessed by receiver operating characteristic (ROC) analysis on C_trough. The area under the ROC curve (AUC_ROC) was used to identify the optimal C_trough threshold. Single vs multiple C_trough at steady state was also evaluated. After a single risperidone C_trough measurement, the AUC_ROC (95% CI) was estimated to be 0.71 (0.69-0.72) and the optimal C_trough threshold accounting for the lowest number of adherent and non-adherent misclassifications was estimated to be 11.9 ng/mL. After multiple C_trough measurements, the AUC_ROC (95% CI) increased up to 0.85 (0.84-0.87) for three C_trough measurements. The optimal probability threshold to reliably discriminate between adherent and non-adherent patients was estimated to be 0.51. Using this model which is reflective of typical adherence to antipsychotic medication, we found that three consecutive steady-state C_trough measurements are needed for an accurate and precise diagnostic test to discriminate between patients who are adherent or non-adherent to treatment.

Population pharmacokinetic model of lithium and drug compliance assessment

Population pharmacokinetic analysis of lithium during therapeutic drug monitoring and drug compliance assessment was performed in 54 patients and 246 plasma concentrations levels were included in this study. Patients received several treatment cycles (1-9) and one plasma concentration measurement for each patient was obtained always before starting next cycle (pre-dose) at steady state. Data were analysed using the population approach with NONMEM version 7.2. Lithium measurements were described using a two-compartment model (CL/F=0.41Lh^-1, V₁/F=15.3L, Q/F=0.61Lh^-1, and V₂/F = 15.8L) and the most significant covariate on lithium CL was found to be creatinine clearance (reference model). Lithium compliance was analysed using inter-occasion variability or Markovian features (previous lithium measurement as ordered categorical covariate) on bioavailability parameter. Markov-type model predicted the lithium compliance in the next cycle with higher success rate (79.8%) compared to IOV model (65.2%) and reference model (43.2%). This model becomes an efficient tool, not only being able to adequately describe the observed outcome, but also to predict the individual drug compliance in the next cycle. Therefore, Bipolar disorder patients can be classified regarding their probability to become extensive or poor compliers in the next cycle and then, individual probabilities lower than 0.5 highlight the need of intensive monitoring, as well as other pharmaceutical care measurements that might be applied to enhance drug compliance for a better and safer lithium treatment.