A survival mediation model with Bayesian model averaging

Radiation-Induced Lymphopenia is a Causal Mediator of Survival After Chemoradiation Therapy for Esophagus Cancer

Purpose

Radiation-induced lymphopenia (RIL) is common during chemoradiation therapy. Severe lymphopenia is associated with reduced survival. Proton beam therapy (PBT), with its substantially more compact dose distributions, spares circulating lymphocytes and immune organs at risk to a greater extent than photon therapy. Recent studies comparing PBT to photon radiation therapy, specifically intensity-modulated radiation therapy (IMRT) for esophageal cancer (EC), showed that the incidence of grade 4 RIL (G4RIL) is significantly reduced among patients receiving PBT for EC. However, whether the extent of this reduction has a direct causative link with improved survival is unknown. This study applies causal mediation analysis to answer this question.

Methods and Materials

We retrospectively assessed 734 patients treated with concurrent chemoradiation therapy for biopsy-proven EC from 2004 to 2017. To address the potential for bias in the choice of radiation modality, propensity score analysis was used to evaluate and reduce imbalances between the PBT and IMRT cohorts. Causal mediation analysis was applied to decompose the total effect of radiation modality on overall survival (OS) into indirect (mediated through G4RIL) and direct effects.

Results

We found that PBT was associated with a significantly lower incidence of G4RIL and prolonged OS compared with IMRT (odds ratio, 0.41; 95% CI, 0.28-0.60; P < .001). In the propensity-matched cohort of 506 patients (253 PBT, 253 IMRT), G4RIL risk reduction with PBT versus IMRT translated into a 5% reduction in the relative rate of death (P = .032). Mediation of G4RIL explained ∼14.5% of the difference in OS.

Conclusions

G4RIL was found to mediate survival; however, a statistically significant direct effect of PBT on survival was not observed. In other words, the statistical significance of survival benefit from protons over photons in this EC cohort was lost in the absence of G4RIL risk reduction.

Flexible evaluation of surrogate markers with Bayesian model averaging

When long-term follow up is required for a primary endpoint in a randomized clinical trial, a valid surrogate marker can help to estimate the treatment effect and accelerate the decision process. Several model-based methods have been developed to evaluate the proportion of the treatment effect that is explained by the treatment effect on the surrogate marker. More recently, a nonparametric approach has been proposed allowing for more flexibility by avoiding the restrictive parametric model assumptions required in the model-based methods. While the model-based approaches suffer from potential mis-specification of the models, the nonparametric method fails to give desirable estimates when the sample size is small, or when the range of the data does not follow certain conditions. In this paper, we propose a Bayesian model averaging approach to estimate the proportion of treatment effect explained by the surrogate marker. Our procedure offers a compromise between the model-based approach and the nonparametric approach by introducing model flexibility via averaging over several candidate models and maintains the strength of parametric models with respect to inference. We compare our approach with previous model-based methods and the nonparametric method. Simulation studies demonstrate the advantage of our method when surrogate supports are inconsistent and sample sizes are small. We illustrate our method using data from the Diabetes Prevention Program study to examine hemoglobin A1c as a surrogate marker for fasting glucose.

Landmark mediation survival analysis using longitudinal surrogate

Clinical cancer trials are designed to collect radiographic measurements of each patient's baseline and residual tumor burden at regular intervals over the course of study. For solid tumors, the extent of reduction in tumor size following treatment is used as a measure of a drug's antitumor activity. Statistical estimation of treatment efficacy routinely reduce the longitudinal assessment of tumor burden to a binary outcome describing the presence versus absence of an objective tumor response as defined by RECIST criteria. The objective response rate (ORR) is the predominate method for evaluating an experimental therapy in a single-arm trial. Additionally, ORR is routinely compared against a control therapy in phase III randomized controlled trials. The longitudinal assessments of tumor burden are seldom integrated into a formal statistical model, nor integrated into mediation analysis to characterize the relationships among treatment, residual tumor burden, and survival. This article presents a frameworkfor landmark mediation survival analyses devised to incorporate longitudinal assessment of tumor burden. R ² effect-size measures are developed to quantify the survival treatment mediation effects using longitudinal predictors. Analyses are demonstrated with applications to two colorectal cancer trials. Survival prediction is compared in the presence versus absence of longitudinal analysis. Simulation studies elucidate settings wherein patterns of tumor burden dynamics require longitudinal analysis.

Radiomics, deep learning and early diagnosis in oncology

Medical imaging, including X-ray, computed tomography (CT), and magnetic resonance imaging (MRI), plays a critical role in early detection, diagnosis, and treatment response prediction of cancer. To ease radiologists' task and help with challenging cases, computer-aided diagnosis has been developing rapidly in the past decade, pioneered by radiomics early on, and more recently, driven by deep learning. In this mini-review, I use breast cancer as an example and review how medical imaging and its quantitative modeling, including radiomics and deep learning, have improved the early detection and treatment response prediction of breast cancer. I also outline what radiomics and deep learning share in common and how they differ in terms of modeling procedure, sample size requirement, and computational implementation. Finally, I discuss the challenges and efforts entailed to integrate deep learning models and software in clinical practice.

Predicting outcomes of phase III oncology trials with Bayesian mediation modeling of tumor response

Pivotal cancer trials often fail to yield evidence in support of new therapies thought to offer promising alternatives to standards-of-care. Conducting randomized controlled trials in oncology tends to be considerably more expensive than studies of other diseases with comparable sample size. Moreover, phase III trial design often takes place with a paucity of survival data for experimental therapies. Experts have explained the failures on the basis of design flaws which produce studies with unrealistic expectations. This article presents a framework for predicting outcomes of phase III oncology trials using Bayesian mediation models. Predictions, which arise from interim analyses, derive from multivariate modeling of the relationships among treatment, tumor response, and their conjoint effects on survival. Acting as a safeguard against inaccurate pre-trial design assumptions, the methodology may better facilitate rapid closure of negative studies. Additionally the models can be used to inform re-estimations of sample size for under-powered trials that demonstrate survival benefit via tumor response mediation. The methods are applied to predict the outcomes of two colorectal cancer studies. Simulation is used to evaluate and compare models in the absence versus presence of reliable surrogate markers of survival.