Waterfall plot(s) to display the results in oncology and non-oncology clinical trials

The waterfall plot(s) started as a visualization tool in oncology studies to display the results of tumor burden, tumor size (or change in tumor size), the tumor responses have gained popularity and appeared in many publications. The application of the waterfall plot has gone beyond the oncology clinical trials. 

According to a paper by Gillespie (2012) Understanding Waterfall Plots,

Waterfall plots are graphic illustrations of data that can vary from audio frequencies to clinical trial patient information and results. In oncology, for example, a waterfall plot may be used to present each individual patient’s response to a particular drug based on a parameter, such as tumor burden. The horizontal (x) axis across the plot may serve as a baseline measure; vertical bars are drawn for each patient, either above or below the baseline. The vertical (y) axis may be used to measure maximum percent change from baseline, e.g., percent growth or reduction of the tumor by radiologic measurement. Those vertical bars that are above the line represent nonresponders or progressive disease. Vertical bars below the baseline (x) axis are drawn for each patient that has achieved some degree of tumor reduction, often depicted as negative percent.

In general, waterfall plots go from the worst value, such as greatest progression of disease, on the left side of the plot, to the best value, i.e., most reduction of tumor, on the right side of the plot; this can also be shown by shifting the graph to a similar presentation, moving from the worst outcomes on the bottom to the best outcomes on the top. The length of each vertical bar hanging below the horizontal axis increases as the plot moves to the right side of the graph, thus resembling a waterfall and giving the graph its name. Thus, the data are not presented randomly, or in order of when a patient first enrolled in a trial, but are organized in order to provide a clear picture of the study population’s results: from worst to best, based on the parameters included. 

The waterfall plot(s) has the following features: 

• It’s basically a bar graph, where each bar typically represents a patient; they are usually ordered from worst results to best.
• The horizontal axis is generally chosen to be a baseline measure, and the bars may go either above or below the baseline. 
• The x-axis is generally the subject number. If the x-axis is not labeled, it defaults to be the subject number. The subjects are listed according to the rank from worst results (on the left) to best results (on the right)
• The y-axis is generally used to quantify response to treatment; for instance, it might represent the percent of growth or reduction in a tumor while a patient is undergoing radiology. Negative bars would show reduction; positive bars would be patients whose cancer is still progressing or non-responders.
• For a study with multiple arms, each arm will have its own waterfall plot. For a study with three treatment arms, there will be three waterfall plots. The difference can be seen by comparing the patterns from different waterfall plots. 

In Advani (2018) CD47 Blockade by Hu5F9-G4 and Rituximab in Non-Hodgkin’s Lymphoma, a waterfall plot was used to display the change in tumor-lesion size with treatments of 5F9 and Rituximab. The waterfall plot showed the best overall change in the size of tumor target lesions among patients with diffuse large B-cell lymphoma (DLBCL; indicated by an asterisk) or follicular lymphoma, according to the maintenance dose received. The y-axis is the percentage changes in the tumor burden of target lesions and the x-axis is the patient number.

In Kopetz et al (2019) Encorafenib, Binimetinib, and Cetuximabin BRAF V600E–Mutated Colorectal Cancer, three waterfall plots were used to display the differences in patterns in best percentage change in the size of target lesions among three treatment groups (triple-therapy, double-therapy, and control groups). Notice that each treatment group has its own waterfall plot. Y-axis is the best percentage change from baseline in tumor size of target lesion. The X-axis is the subject number (even though it is not labeled). 

Waterfall plot(s) has been used in studies beyond the oncology studies. Here are some examples:

In Vichinsky et al (2019) A Phase 3 Randomized Trial of Voxelotor in Sickle Cell Disease, three waterfall plots were used to display the treatment effect in change in hemoglobin level of Vexelotor comparing to Placebo. The y-axis is the change in hemoglobin level from baseline to week 24 (g/dL) and the x-axis is the subject number for each treatment group (even though it is not labeled). 

In Nathan et al (2020) Efficacy of Pirfenidone in the Context of Multiple Disease Progression Events in Patients With Idiopathic Pulmonary Fibrosis, two colorful waterfall plots (one for pirfenidone group and one for the placebo group) were used to display the pattern and distribution of frequency and type of adverse outcome (or disease progression) events including the decline in 6MWD, the decline in %FVC, respiratory-related hospitalization, death, and combination of them. The y-axis is the number of events and the x-axis is the patient number for each treatment group. 

In a retrospective pretest-posttest study with no controls by Sanchez et al (2019) Multiple lifestyle interventions reverses hypertension, two waterfall plots (one for SBP and one for DBP) were used to display the pre-post change in systolic and diastolic blood pressure to indicate the NEWSTART Lifestyle intervention was an effective and rapid means to decrease SBP and DBP.

While waterfall plots can visually show the treatment effects either change from baseline or between treatment groups, there are drawbacks as well. 

According to Kim et al (2019) Assessment of Accuracy of Waterfall Plot Representations of Response Rates in Cancer Treatment Published in Medical Journals, the article assessed 126 studies published in 6 journals where waterfall plots were used to show visual response rates. The author concludes that that waterfall plots are used more frequently over time and exaggerate the visual estimate of the response rate.

In a paper by Shao et al Use and Misuse of Waterfall Plots, the authors concluded that "there was substantial variability in criteria used to generate published waterfall plots. Waterfall plots are subject to substantial variability in criteria used to define them and are influenced by measurement errors; they should be generated by trained radiologists. Caution should be exercised when interpreting the results of waterfall plots in the context of clinical trials."

Waterfall plots can be generated in SAS. There are quite some papers discussing the tips and tricks in generating waterfall plots: 

• Create a waterfall plot in SAS
• Waterfall plot: two different approaches, one beautiful graph 
• Graphical Results in Clinical Oncology Studies
• Useful SAS techniques in Efficacy Analysis for Oncology studies
• Plotting Against Cancer: Creating Oncology Plots Using SAS®