Graphical Presentation of Patient-Treatment Interaction Elucidated by Continuous Biomarkers

Abstract

Summary Background: Biomarkers providing evidence for patient-treatment interaction are key in the development and practice of personalized medicine. Knowledge that a patient with a specific feature – as demonstrated through a biomarker – would have an advantage under a given treatment vs. a competing treatment can aid immensely in medical decision-making. Statistical strategies to establish evidence of continuous biomarkers are complex and their formal results are thus not easy to communicate. Good graphical representations would help to translate such findings for use in the clinical community. Although general guidelines on how to present figures in clinical reports are available, there remains little guidance for figures elucidating the role of continuous biomarkers in patient-treatment interaction (CBPTI). Objectives: To combat the current lack of comprehensive reviews or adequate guides on graphical presentation within this topic, our study proposes presentation principles for CBPTI plots. In order to understand current practice, we review the development of CBPTI methodology and how CBPTI plots are currently used in clinical research. Methods: The quality of a CBPTI plot is determined by how well the presentation provides key information for clinical decision-making. Several criteria for a good CBPTI plot are proposed, including general principles of visual display, use of units presenting absolute outcome measures, appropriate quantification of statistical uncertainty, correct display of benchmarks, and informative content for answering clinical questions especially on the quantitative advantage for an individual patient with regard to a specific treatment. We examined the development of CBPTI methodology from the years 2000-2014, and reviewed how CBPTI plots were currently used in clinical research in six major clinical journals from 2013-2014 using the principle of theoretical saturation. Each CBPTI plot found was assessed for appropriateness of its presentation and clinical utility. Results: In our review, a total of seven methodological papers and five clinical reports used CBPTI plots which we categorized into four types: those that distinguish the outcome effect for each treatment group; those that show the outcome differences between treatment groups (by either partitioning all individuals into subpopulations or modelling the functional form of the interaction); those that evaluate the proportion of population impact of the biomarker; and those that show the classification accuracy of the biomarker. The current practice of utilizing CBPTI plots in clinical reports suffers from methodological shortcomings: the lack of presentation of statistical uncertainty, the outcome measure scaled by relative unit instead of absolute unit, incorrect use of benchmarks, and being non-informative in answering clinical questions. Conclusions: There is considerable scope for improvement in the graphical representation of CBPTI in clinical reports. The current challenge is to develop instruments for high-quality graphical plots which not only convey quantitative concepts to readers with limited statistical knowledge, but also facilitate medical decision-making.

Confidence interval estimation for the changepoint of treatment stratification in the presence of a qualitative covariate‐treatment interaction

“…The results are shown in Table . In Figure , the estimated hazard ratio between the treatment groups in dependence of age is illustrated and a pointwise 95% confidence interval, as described in Shen et al, is given (see supplemental material, Section S1.2).…”

“…When a qualitative covariate‐treatment interaction is observed, the covariate value that is associated with no difference between the two treatment groups can be estimated from the regression coefficients for treatment and the covariate‐treatment interaction. This is the value of the covariate, where the difference between the mean outcomes of the treatments is zero or the odds ratio or the hazard ratio between the treatments is one and the direction of the treatment effect changes . We call the covariate value, where the superior treatment changes, the changepoint of treatment stratification in this article.…”

“…This is the value of the covariate, where the difference between the mean outcomes of the treatments is zero or the odds ratio or the hazard ratio between the treatments is one and the direction of the treatment effect changes. 12 We call the covariate value, where the superior treatment changes, the changepoint of treatment stratification in this article. As this estimate relies on a finite sample and is prone to sampling variability, a confidence interval for that value should be estimated and provided to indicate uncertainty, as commonly recommended for treatment effect estimates.…”

Constructing treatment selection rules based on an estimated treatment effect function: different approaches to take stochastic uncertainty into account have a substantial effect on performance

“…[4]) as well as both together (e.g. [7, 8]) have been suggested. Mackey and Bengtsson, Riddell et al [1, 3] suggest to construct a confidence interval for the root of θ ( x ) (with respect to 0 or another threshold), and similarly [2] suggest to compute horizontal confidence intervals.…”

Visualizing the (Causal) Effect of a Continuous Variable on a Time-To-Event Outcome

“…5 Graphics designed for displaying the effect of a continuous covariate on a time-to-event outcome exist but are rarely used in practice. [10][11][12][13][14][15] Most of these approaches fail to represent the treatment effect over time and as a function of the continuous variable simultaneously. Some rely on summary statistics, such as landmark survival probabilities or restricted mean survival times, 12,13 whereas others rely on rather complex estimation strategies that may not use easily interpretable absolute measures of effect.…”