In current clinical trial development, historical information is receiving more attention as it provides utility beyond sample size calculation. Meta-analytic-predictive (MAP) priors and robust MAP priors have been proposed for prospectively borrowing historical data on a single endpoint. To simultaneously synthesize control information from multiple endpoints in confirmatory clinical trials, we propose to approximate posterior probabilities from a Bayesian hierarchical model and estimate critical values by deep learning to construct pre-specified strategies for hypothesis testing. This feature is important to ensure study integrity by establishing prospective decision functions before the trial conduct. Simulations are performed to show that our method properly controls family-wise error rate and preserves power as compared with a typical practice of choosing constant critical values given a subset of null space. Satisfactory performance under prior-data conflict is also demonstrated. We further illustrate our method using a case study in Immunology.
Introduction: We evaluate safety and efficacy of risankizumab (RZB) vs. placebo (PBO) in adult patients with moderate-to-severe non-pustular palmoplantar psoriasis (PPPsO).
Method: IMMprint, a Ph3b study evaluated safety and efficacy of RZB vs. PBO in moderate-to-severe PPPsO patients with a static physician’s global assessment(sPGA) of moderate or severe ≥3, palmoplantar psoriasis area and severity index(PPASI) ≥8 and at least 1 additional PsO plaque. The 52-week treatment was split into period A, patients randomized (1:1) to RZB 150-mg or PBO, and period B, patients continuing RZB or switching from PBO to RZB (PBO/RZB). Primary endpoint was achievement of palmoplantar investigator’s global assessment (ppIGA) 0/1 with at least 2-point reduction from baseline at Wk16. Ranked secondary endpoints include ≥75% & 90% improvement in PPASI (PPASI75, PPASI90), sPGA 0/1 and 100% improvement in PPASI(PPASI100) at Wk16.
Results: Of 174 enrolled, 87(mean age[SD]: 56.9[12.9] years) were randomized to RZB and 87(mean age[SD]: 53.9[14.3] years) were randomized to PBO. Baseline characteristics were similar except for numerical difference in PsA (RZB vs. PBO: 11.5% vs 4.6%, respectively). At Wk16, a significantly higher proportion receiving RZB achieved ppIGA 0/1 than PBO(33.3% vs. 16.1%, p = 0.006). Patients receiving RZB also demonstrated significantly higher responses than patients receiving PBO in all ranked secondary endpoints: PPASI75 (42.5% vs 14.9%, P < 0.001); PPASI90 (27.6% vs 5.7%, p<0.001); sPGA 0/1 (32.2% vs 11.5%, p < 0.001); and PPASI100 (17.2% vs. 1.1%, p < 0.001). Four patients (3 RZB: 1 PBO) discontinued drug in period A. At Wk52, ppIGA 0/1 was achieved by 50.6% of RZB patients and 61.7% of PBO/RZB group. Proportion of patients achieving PPASI75 at Wk52 was 57.5%(RZB) vs. 65.4% (PBO/RZB). Achievement of sPGA 0/1 (%RZB vs. %PBO/RZB) was 43.7% vs. 66.7%; PPASI100 was achieved by 26.4% (RZB) and 37.0% (PBO/RZB) patients at Wk52. Proportion of patients with adverse events were 29.1%(RZB) and 23.0%(PBO) in period A, and 49.4%(RZB) and 35.8% (PBO/RZB) in period B. One RZB patient with prior cardiovascular risk factors had an adjudicated myocardial infarction and subsequently died after 140-day follow-up period. The number of patients with COVID-19 were 3 (RZB; 1 serious infection) and 2(PBO) in period A and 11 (RZB) and 7 (PBO/RZB) in period B.
Conclusion: This study demonstrates RZB can provide effective improvement compared to PBO by Wk16 with continuous improvement until Wk52 with no new safety signals in patients with moderate-to-severe PPPsO.
Deep learning is a subfield of machine learning used to learn representations of data by successive layers. Remarkable achievements and breakthroughs have been made in image classification, speech recognition, et cetera, but the full capability of deep learning is still under exploration. As statistical researchers and practitioners, we are especially interested in leveraging and advancing deep learning techniques to address important and impactive problems in biomedical and other related fields. In this article, we provide a basic introduction to Feedforward Neural Networks (FNN) along with some intuitive explanations behind its strong functional representation. Guidance is provided on how to choose quite a few hyperparameters in neural networks for a specific problem. We further discuss several more advanced frameworks in deep learning. Some successful applications of deep learning in biomedical fields are also demonstrated. With this beginner's guide, we hope that interested readers can include deep learning in their toolbox to tackle future real-world questions and challenges.
As one of the most important estimators in classical statistics, the uniformly minimum variance unbiased estimator (UMVUE) has been adopted for point estimation in many statistical studies, especially for small sample problems. Moving beyond typical settings in the exponential distribution family, it is usually challenging to prove the existence and further construct such UMVUE in finite samples. For example in the ongoing Adaptive COVID-19 Treatment Trial (ACTT), it is hard to characterize the complete sufficient statistics of the underlying treatment effect due to pre-planned modifications to design aspects based on accumulated unblinded data. As an alternative solution, we propose a Deep Neural Networks (DNN) guided ensemble learning framework to construct an improved estimator from existing ones. We show that our estimator is consistent and asymptotically reaches the minimal variance within the class of linearly combined estimators. Simulation studies are further performed to demonstrate that our proposed estimator has considerable finite-sample efficiency gain. In the ACTT on COVID-19 as an important application, our method essentially contributes to a more ethical and efficient adaptive clinical trial with fewer patients enrolled.
In confirmatory clinical trials, it has been proposed to use a simple iterative graphical approach to construct and perform intersection hypotheses tests with a weighted Bonferroni-type procedure to control Type I errors in the strong sense. Given Phase II study results or other prior knowledge, it is usually of main interest to find the optimal graph that maximizes a certain objective function in a future Phase III study. In this article, we evaluate the performance of two existing derivative-free constrained methods, and further propose a deep learning enhanced optimization framework. Our method numerically approximates the objective function via feedforward neural networks (FNNs) and then performs optimization with available gradient information. It can be constrained so that some features of the testing procedure are held fixed while optimizing over other features. Simulation studies show that our FNN-based approach has a better balance between robustness and time efficiency than some existing derivative-free constrained optimization algorithms. Compared to the traditional stochastic search method, our optimizer has moderate multiplicity adjusted power gain when the number of hypotheses is relatively large. We further apply it to a case study to illustrate how to optimize a multiple testing procedure with respect to a specific study objective.
Design, VLSI implementation, and experimental validation of a resource-optimized machine-learning algorithm for epilepsy seizure detection is presented. The algorithm uses only signals from the frontal and the front-temporal lobes EEG electrodes while yielding a seizure detection performance competitive to the standard full EEG systems. The experimental validations prove the possibility of conducting accurate seizure detection using quickly-mountable dry-electrode headsets without the need for uncomfortable/painful through-hair electrodes or adhesive material. The compact VLSI implementation of the algorithm is also presented and resource optimization techniques are discussed. The optimized implementation is uploaded on an Actel Igloo AGL250 low-power FPGA, requires 1237 logic elements, consumes 110μW dynamic power, and yields a detection latency of 10.2μs. The measurement results from the FPGA implementation on data from 23 patients (198 seizures in total) shows a seizure detection sensitivity and specificity of 92.5% and 80.1%, respectively.
In recent pharmaceutical drug development, adaptive designs of clinical trials have become more and more appealing due to ethical considerations and the ability to accommodate uncertainty while conducting the trial. In addition to design optimization based on certain criteria, it is also meaningful to identify a better hypothesis testing strategy for a specific design. In this book chapter, we review a deep learning–based two-stage approach to construct test statistics and corresponding critical values. This method is applied to a response adaptive randomization design and a sample size reassessment design to gain power in hypothesis testing. Discussions, including future works, are also provided.
'%3e%3cpath fill='%23012169' d='M0 0v30h60V0z'/%3e%3cpath stroke='%23fff' stroke-width='6' d='m0 0 60 30m0-30L0 30'/%3e%3cpath stroke='%23C8102E' stroke-width='4' d='m0 0 60 30m0-30L0 30' clip-path='url(%23b)'/%3e%3cpath stroke='%23fff' stroke-width='10' d='M30 0v30M0 15h60'/%3e%3cpath stroke='%23C8102E' stroke-width='6' d='M30 0v30M0 15h60'/%3e%3c/g%3e%3c/svg%3e)