Thursday, Aug 7: 10:30 AM - 12:20 PM
4225
Contributed Papers
Music City Center
Room: CC-101D
Main Sponsor
Biopharmaceutical Section
Presentations
Accurate prediction of key clinical trial milestones is crucial for efficient trial planning, particularly in event-driven studies. We propose a flexible framework integrating enrollment prediction, time-to-event modeling, and dropout estimation to project milestone events such as interim and final analyses. Our framework leverages Bayesian methodology for enrollment prediction, using a nonhomogeneous Poisson process with a quadratic time-varying rate function to model accrual dynamics. For event prediction, we incorporate piecewise exponential models with breakpoint estimation, enabling flexible hazard rate assumptions, alongside parametric mixture cure rate models, including exponential, Weibull, and Gompertz mixture cure models, to account for long-term survivors and non-proportional hazards. Additionally, we develop advanced time-to-dropout models with various distributional assumptions. Our simulation-based approach enhances analysis time estimation and outperforms conventional methods. Retrospective validation demonstrates substantial improvements in prediction accuracy. To support implementation, we developed an R Shiny application with an intuitive user interface.
Keywords
Bayesian modeling
Event prediction
Mixture cure models
Clinical trial
Enrollment forecasting
Traditional clinical trial monitoring relies heavily on manual site visits and source data verification (SDV), making the process both time-consuming and costly. We introduce a machine learning (ML) and deep learning (DL) enabled system that automates real-time risk assessment based on risk-based monitoring (RBM) and quality tolerance limits (QTL) principles. By integrating high-dimensional, diverse data streams and types, our approach proactively identifies and predicts potential issues that could compromise patient safety and data integrity. Unlike static, threshold-based methods, our models adapt dynamically to accumulating trial data, reducing manual SDV and streamlining resource allocation. Preliminary findings suggest accelerated detection of high-risk sites and data anomalies, improved patient safety outcomes, and significant operational efficiencies.
Keywords
Clinical trial monitoring
Quality tolerance limits
Machine learning
Deep learning
Patient safety
Data integrity
A suite of complementary approaches are used across the data life cycle at patient, site, trial, portfolio levels to ensure the quality and data integrity of our clinical trials. These include elements of trial design, process, technology, analytics and automation. One important element is remote/centralized risk-based monitoring. This involves identification of critical data and processes, risk indicators and thresholds, and development and implementation of a monitoring plan to direct oversight efforts informed by known or emerging areas of risk.
This presentation is focused on using integrated data, statistical analytics and automation in the context of remote Risk Based Monitoring (rRBM) to increase the effectiveness of on-site monitoring by Clinical Research Associates (CRAs) of clinical sites participating in our trials. The objective is to enable more targeted monitoring that is informed and triggered by indicators of risk. The centralized approach using integrated data is expected to be more sensitive for earlier detection of sites with outlying performance and/or data, that may indicate issues with site conduct.
Keywords
remote Risk Based Monitoring (rRBM)
Clinical Trial Oversight
earlier detection of site issue
Covariate adjustment in randomized trials remains underutilized despite its potential to improve precision while requiring the same or weaker assumptions for validity compared to an unadjusted analysis. Such methods may not be directly compatible with group sequential designs (GSDs), which are commonly used for pre-planned interim analyses.
'impart' is an R package that allows users to design, monitor, and analyze randomized trials which can incorporate both covariate adjustment and pre-planned interim analyses. The package includes functions for pre-trial planning, monitoring ongoing studies, and performing analyses. Functions facilitate planning trials with continuous, binary, and ordinal outcomes. Monitoring and analyses can be done using pre-specified methods (e.g., G-computation), or user-supplied functions. Variance estimates are computed using the nonparametric bootstrap and can be orthogonalized to enforce the independent increments assumption necessary to utilize GSD stopping boundaries. 'impart' also facilitates information monitoring (i.e. continuous sample size re-estimation), allowing recruitment to be adapted to accruing data, avoiding under- or overpowered trials.
Keywords
Randomized Trials
Covariate Adjustment
Adaptive Designs
Sample Size Re-estimation
Information Monitoring
Group Sequential Designs
Co-Author(s)
Kelly Van Lancker, Ghent University
Michael Rosenblum, Johns Hopkins University, Bloomberg School of Public Health
First Author
Joshua Betz, Johns Hopkins Bloomberg School of Public Health
Presenting Author
Joshua Betz, Johns Hopkins Bloomberg School of Public Health
Evidence-based quantitative decision-making is crucial at every stage of clinical development. Sponsors often look at ongoing clinical trial data either blindly or unblindly to make critical decisions. In this presentation, we focus on the discussion of interim analysis based on blinded data. Blinded data analysis is commonly performed during clinical trials to monitor the data quality, estimate the variation of currently observed data, or prepare the scenario planning before the final study readout. However, given the nature of the blindness, sometimes, the value of blinded analysis based on ongoing trial data could be limited for decision making. In this presentation, quantitative and visualization techniques to maximize the value of blinded analysis and support effective communication with senior management based on blinded data under quantified uncertainty are presented through a real clinical trial case study.
Keywords
blinded data analysis
clinical trials
scenario planning
Randomized double-blind clinical trials are considered the "gold standard" for evaluating the effects of a new intervention. Traditionally, only the treatment assignment is blinded to the investigator and sponsor. However, differences in response variables between the investigational and control arms is typically hypothesized. If such differences are large, awareness of the response variables could reveal the treatment assignment and potentially introduce bias into the study. To mitigate this, responses are sometimes blinded to prevent reverse guessing, but this complicates study monitoring. However, are such measures really required to protect study integrity? We propose an unblinding risk score (URS) to quantify the risk that blinded site or Sponsor personnel could correctly guess the treatment assignment based on an individual patient's efficacy response. Our findings are applied to several published studies to demonstrate how the URS can evaluate unblinding risk, and we demonstrate an interactive tool built in R Shiny tool to help study teams assess this risk before deciding to mask or not mask the efficacy data.
Keywords
Clinical trials
data masking
unblinding
risk assessment
randomized double-blinded placebo-controlled studies
unblinding risk