What is the value of adaptive designs? Estimating expected value of sample information for adaptive trial designs.

目的：开发样本信息期望值(EVSI)方法，允许有效比较传统固定方法和适应性方法在试验设计中的经济价值。背景：卫生保健决策者，如英国国家健康与护理卓越研究所(NICE)，通过成本效益分析来指导他们的决策。这些是对疾病干预的成本和效果进行的基于试验或模型的比较。消除这些决策建议中不确定性的价值是完美信息的期望值(EVPI)。仅消除一部分不确定性的价值是部分完全信息的期望值(EVPPI)。最后，EVSI是通过特定的试验设计减少而不是消除不确定性的经济价值。适应性设计是那些随着结果的积累而适应的设计。例子包括：放弃或增加治疗武器的多臂多阶段(MAMS)试验；改变随机化比率的试验；以及将招募重点放在最有可能受益或受益最不确定的患者上的人口充实试验。这种设计可以达到相同的准确度和精密度，但分配给不良或有害治疗的患者更少，总体样本量更小，持续时间更短，或试验成本更低。适应性设计还可以检验多个假设，潜在地提供更好的投资回报。适应性设计越来越受欢迎，但仍是例外，而不是规则。EVSI仅应用于有限范围的自适应设计，计算负担是更广泛使用的主要障碍。我将把计算效率高的EVSI方法扩展和开发到适应性设计中，以评估它们与传统设计相比的经济价值。工作流程1.将有效的EVSI估计方法扩展到适应性设计。当成本效益分析依赖于复杂的模型时，例如马尔可夫多状态模型，估计任何试验的EVSI都可能是一个巨大的计算挑战。对于适应性设计来说，情况尤其如此。最近的工作集中在成本-效果模型的近似方法上，包括通过灵活的高斯过程或广义加性模型的回归，通过样条或泰勒级数的近似，以及成本和效果分布的矩匹配。这些都是为了评估固定的设计，而不是适应性设计。我将把这些方法扩展到评估不同自适应设计的经济效益。工作流2.开发改进的蒙特卡罗抽样方案来估计自适应设计的EVSI。我将研究的第二种方法是高效抽样方案。EVSI的估计依赖于随机数生成的嵌套循环，称为蒙特卡罗模拟。暴力蒙特卡罗抽样通常太慢，不适用于EVSI。先进的蒙特卡罗方案可以减少必要的模拟次数。其中包括多水平蒙特卡罗，它用一个较低的方差替代估计目标，以及准蒙特卡罗，它使用准随机数而不是随机数来减少方差。这些已经应用于EVPPI，但不适用于EVSI；我将把它们扩展到EVSI，并进一步扩展到适应性设计。工作流程3.应用于真实世界的成本-效果模型和适应性试验设计。实际应用包括抑郁症治疗的评估，房颤中预防中风的抗凝剂，以及髋关节置换假体的比较。由于大量可用的假体，MAMS试验可能被考虑用于髋关节置换术，而由于对轻度抑郁症患者的治疗效果不确定，人群丰富可能被应用于抑郁症的治疗。将为这些试验设计中最有前途的设计提交资金申请。

Aim: To develop expected value of sample information (EVSI) methods that allow efficient comparison of economic value of conventional fixed and adaptive approaches for a trial design.Background: Healthcare decision makers, such as the National Institute for Health and Care Excellence (NICE) in the UK, are guided in their decisions by cost-effectiveness analysis. These are trial- or model-based comparisons of the costs and effects of interventions for diseases. The value of eliminating the uncertainty in these decision recommendations is the expected value of perfect information (EVPI). The value of eliminating only a subset of the uncertainties is the expected value of partial perfect information (EVPPI). Finally, the EVSI is the economic value of reducing, rather than eliminating, uncertainty via a particular trial design.Adaptive designs are those that adapt in response to results as they accumulate. Examples include: multi-arm multi-stage (MAMS) trials that drop or add treatment arms; trials that change randomisation ratios; and population enrichment trials that focus recruitment on patients most likely to benefit, or in whom benefit is most uncertain. Such designs can achieve the same level of accuracy and precision but with fewer patients assigned to non-performing or harmful treatments, smaller overall sample sizes, shorter durations, or lower trial costs. Adaptive designs can also test multiple hypotheses, potentially offering a better return on investment. Adaptive designs have been growing in popularity but remain the exception rather than the rule. EVSI has only been applied to a limited range of adaptive designs, with computational burden being a primary barrier to wider usage. I will extend and develop computationally efficient EVSI methods to adaptive designs to assess their economic value compared with conventional designs.Work stream 1. Extending efficient methods for estimating EVSI to adaptive designs. When cost-effectiveness analysis relies on complex models, such as Markov multi-state models, estimating EVSI for any trial can represent a substantial computational challenge. This is especially true for adaptive designs. Recent work has focussed on methods to approximate the cost-effectiveness model, including regression via flexible Gaussian processes or generalised additive models, approximation via splines or Taylor series, and moment matching of cost and effect distributions. These have been explored for valuing fixed but not adaptive designs. I will extend these methods to valuing the economic benefits of different adaptive designs.Work stream 2. Develop advanced Monte Carlo sampling schemes to estimate the EVSI of adaptive designs.The second approach I will investigate is that of efficient sampling schemes. Estimation of EVSI relies on nested loops of random number generation, called Monte-Carlo simulation. Brute force Monte-Carlo sampling is usually too slow to be practical for EVSI. Advanced Monte-Carlo schemes that reduce the number of necessary simulations are available. These include Multilevel Monte-Carlo, which replaces the estimation target with a lower variance alternative, and Quasi Monte-Carlo, which uses quasi-random rather than random numbers to reduce variance. These have been applied to EVPPI but not EVSI; I will extend them to EVSI and furthermore to adaptive designs. Work stream 3. Application to real world cost-effectiveness models and adaptive trial designs.Practical applications include the evaluation of treatments for depression, anticoagulants for prevention of stroke in atrial fibrillation, and the comparison of prosthetics for hip replacement. MAMS trials may be considered for hip replacement due to the large number of available prosthetics and population enrichment may be applied to depression treatment due to uncertain treatment effects in low severity depression patients. Funding applications will be submitted for the most promising of these trial designs.