Journal of Biopharmaceutical Statistics最新文献

The use of real-world data, containing data from historical clinical studies, to construct an external control arm or to augment a small internal control arm in a randomized control trial can lead to significant improvements in the efficiency of the trial, but it may also introduce bias. To mitigate the risk of potential bias arising from the heterogeneity between the external control and the internal control arms, Bayesian dynamic borrowing, which determines the amount of borrowing by similarity between the two data sources, using power prior approaches and covariate adjustment has been introduced. For binary and continuous outcomes, an approach integrating propensity score for covariate adjustment and Bayesian dynamic borrowing using power prior has been proposed. Here, we extend this approach to survival analysis with the hazard ratio as the estimand. We propose a novel approach for estimating the amount of borrowing using the empirical Bayes method based on the log-hazard ratio between external and internal controls. For inference, the approach uses Bayesian bootstrap in combination with the empirical Bayes method, covariate adjustment, and multiple imputation, taking into account all uncertainty. The performance of our approach is examined by a simulation study. As an illustration, we apply the approach to dynamic borrowing of Flatiron real-world data for CheckMate-057 study for advanced non-squamous non-small cell lung cancer. For this application, we apply multiple imputation for missing covariates and propose a computationally efficient algorithm for computing the total variance of the log hazard ratio estimate. The proposed method can be applied to other endpoints in oncology as well as to other disease areas.

Purpose: Alzheimer's disease (AD) is a neurodegenerative disorder characterized by progressive cognitive decline. We proposed a novel latent multimodal deep learning framework to predict AD cognitive status using clinical, neuroimaging, and genetic data.

Methods: Three hundred and twenty-two patients aged between 55 and 92 from the ADNI database were included in the study. Confirmatory Factor Analysis (CFA) was applied to derive the latent scores of AD cognitive impairments as the outcome. A multimodal deep neural network with three modalities, including clinical data, imaging data, and genetic data, was constructed. Attention layers and cross attention layers were added to improve prediction; modality importance scores were calculated for interpretation. Mean Absolute Error (MAE) and Mean Squared Error (MSE) were used to evaluate the model performance.

Results: The CFA demonstrated good fit to the data. The multimodal neural network of clinical and imaging modalities with attention layers was the best predictive model, with an MAE of 0.330 and an MSE of 0.206. Clinical data contributed the most (35%) to the prediction of AD cognitive status.

Conclusion: Our results demonstrated the attention multimodal model's superior performance in predicting the cognitive impairment of AD, introducing attention layers into the model enhanced the prediction performance.

Clinicians have increasingly turned to F-scores to gauge the accuracy of diagnostic tests. However, the dependency of F-scores on the prevalence of the underlying illness poses challenges, especially when prevalence varies across regions or populations, potentially leading to misdiagnoses. To address this issue, this article presents novel post-test diagnostic precision metrics for continuous tests or biomarkers. These metrics are based on the collective areas under the F-score curves across all conceivable prevalence values. Unlike traditional measures, the proposed metrics remain constant regardless of disease prevalence, enabling fair comparisons of different diagnostic tests and biomarkers' abilities in rule-in, rule-out, and overall accuracy. The article also explores the relationship between the proposed metrics and other diagnostic accuracy measures. Numerical illustrations and a real-world breast cancer dataset exemplify the application of the proposed metrics.

Data-driven decision-making is crucial in drug development, with the predictive probability of success (PoS) being a key quantitative tool. PoS estimates the likelihood of success of a future trial based on the same or surrogate endpoint(s) of interest, utilizing information from interim analyses, or completed historical studies. While it has been extensively studied and broadly applied in clinical practice, there is a growing need of a unified approach for PoS that can effectively incorporate information from surrogate endpoints and multiple historical studies. This paper investigates and assesses a unified Bayesian approach for PoS. We first review PoS based on historical data on the same endpoint and then extend it to include information from a surrogate endpoint with a closed-form solution. Additionally, we utilize a Bayesian meta-analytic approach to incorporate data from multiple historical studies. We illustrate the unified approach with examples from oncology and immunology trials and provide an R package "PPoS" for practical implementation. By integrating the assessment of PoS with information from surrogate endpoints and historical studies, we aim to enhance the decision-making process in drug development.

This paper discusses quality principles for Phase I model-based dose escalation design. It emphasizes that a loss function underlying a dose escalation trial estimator can be usefully interpreted as a quantified representation of the ethical assumptions underlying the treatment decisions to be made in the trial. Based on this principle, it discusses additional general quality design principles developers of clinical trial design methods should consider, including the role of continuous loss functions in quality per Taguchi, and per Deming the role of asymmetric loss functions and the importance of understanding the underlying process and its order of operations. It provides a number of model-based dose escalation designs as examples, including the mTPI as an introductory example, the EWOC design, and the CRM and modifications to it. It introduces some foundational scientific underpinnings and principles of quality philosophy, and explains how the principles apply to the examples. It stresses the importance of an engineering process by which a study is designed to meet identified and investigated user requirements.

Augmented randomized clinical trials are a valuable design option for early phase clinical trials. The addition of external controls could, on the one hand, increase precision in treatment effect estimates or reduce the number of required control patients for a randomized trial but may, on the other hand, introduce bias. We build on previous work on augmented trials with one external control data source in time-to-event settings and extend it to multiple control data sources. In a comprehensive simulation study, we evaluate existing and novel analysis options mainly based on Bayesian hierarchical models as well as propensity score analysis. Different sources of bias are investigated including population (observable and unobservable confounders), data collection (assessment schedule, real-world vs. clinical trial data), and time trend as well as different types of data like individual patient data (with or without baseline covariates) or summary data. Our simulation study provides recommendations in terms of choice of estimation method as well as choice of data sources. Explicit incorporation of the above-mentioned sources of bias in a simulation study is relevant as the magnitude of deviation from the ideal setting has a significant impact on all investigated estimation methods.

Background: Time to Benefit (TTB) is a critical metric in clinical practice, reflecting the duration required to achieve therapeutic goals post-treatment. Traditionally, TTB estimation has relied on Bayesian Weibull regression, which, despite its merits, can be computationally intensive. To address this, we propose and evaluate Frequentist methods as efficient alternatives to approximate Bayesian TTB estimation.

Methods: We evaluated three Frequentist methods, parametric delta, Monte Carlo, and nonparametric bootstrap, for TTB estimation, comparing their performance with the Bayesian approach.

Results: Extensive simulations demonstrated that the proposed Frequentist methods outperformed the Bayesian method in efficiency. Real-world data applications further validated these findings, with the Monte Carlo (MC) method exhibiting significantly faster computational speed compared to the nonparametric bootstrap, while the Bayesian method was the least efficient.

Conclusions: The proposed Frequentist methods offer significant advantages to approximate the Bayesian approach for TTB estimation, particularly in efficiency and practicality. The Monte Carlo method, with its median point estimate and percentile confidence intervals, is the recommended choice for its balance of efficacy and expedience.

The assessment and comparison of biomarkers and diagnostic tests using a benefit-risk framework are essential for evaluating both the accuracy of tests and the clinical implications of diagnostic errors. Traditional measures, such as sensitivity and specificity, often do not fully capture the complexities involved in evaluating tests for diseases with multiple subtypes. Many diseases, such as Alzheimer's, are characterized by multiple stages or classes, and in some cases, like cancers, these classes do not follow a specific order, necessitating a more nuanced approach.This paper extends the net benefit approach, traditionally applied to binary diagnostic tests, to address clinical conditions with multiple unordered subtypes using a tree or umbrella ordering framework. We introduce a novel methodology that expands the diagnostic yield table to account for multisubtypes, allowing for a more comprehensive evaluation of diagnostic tests. This approach incorporates decision-making processes based on net benefit, offering additional insights into the criteria for ruling in or ruling out clinical conditions and highlighting the potential adverse consequences of unnecessary diagnostic workups.Through numerical examples, simulations, and real-world data applications, we demonstrate the flexibility and potential advantages of our proposed framework in handling complex disease scenarios. By accommodating multiple subtypes and providing a structured approach to evaluating the net benefit of diagnostic tests, this methodology offers valuable insights for clinical decision-making. The framework's ability to incorporate the specific characteristics of disease subtypes makes it particularly useful in settings where traditional binary classification measures may fall short. This approach could significantly enhance the accuracy of diagnostic evaluations and support more tailored interventions in clinical practice, thereby improving patient outcomes.

Binary meta-analysis studies with rare outcomes frequently include zero or a small number of observations in study groups, creating a sparsity issue with the data. The corrections applied to eliminate the impact of the zero cell counts introduce a bias to the meta-analysis results and potentially distort the inferences about the treatment effect and heterogeneity among the studies. The boundaries of interval estimates become highly biased due to the sparsity of the data. This study proposes two Bayesian random-effects meta-analysis models based on the beta-binomial model with an arc-sine-square-root transformation. The performance of the models in estimating the treatment effect and the in-between study variance is assessed with an extensive Monte Carlo simulation study, and a frequently referred meta-analysis dataset is revisited. The models provide accurate estimates of treatment effect and heterogeneity parameters without a continuity correction. They provide well-calibrated, narrow interval estimates with sufficient coverage of true treatment effect and in-between study variance. They are robust against zero cell counts, very low event probabilities, and unbalanced, skewed data distributions. Recommendations are given for the practical use of the proposed models, and the required model scripts are provided to implement the models using R software.

A cancer trial with an immunotherapy or antibody drug conjugate often has a certain delay/crossing time before the drug to take effect. In this paper, we propose to call the events that occur during and after the delay/crossing time as non-informative events and informative events, respectively. We propose to call the rate of number of informative events divided by total number of events as informative event rate ($\gamma$), though this rate has been used in the literature. We show three innovative usages of $\gamma$ under non-proportional hazards (NPH) setting: (1) based on $\gamma$, the minimum average hazard ratio ($\mathrm{aH}{R}_{\min }$) can be calculated analytically and used to determine whether trials are worth being conducted for a test drug to get a meaningful average hazard ratio (aHR) at the planning stage; (2) based on a series of $\gamma$, aHR and power can be calculated and a proper design can be selected for a trial with a targeted aHR at the design stage; (3) based on $\gamma$, a better interim analysis timing to ensure a certain probability for early efficacy/futility stopping can be determined during the course of a study. aHR and the probability for early efficacy/futility stopping under different enrollment scenarios in a simulation were verified by calculation. We propose the concepts of the informative event rate ($\gamma$), $\mathrm{aH}{R}_{\min }$, and a targeted aHR and use them in study determination, study design, and interim analysis monitoring under an NPH setting with a delay/crossing time.