Waterfall Plot Generator for Oncology Response Data

What is a waterfall plot?

A waterfall plot in oncology is a bar chart with one vertical bar per patient. Each bar shows that patient’s best percentage change in tumor size from baseline, usually the change in the sum of the longest diameters of the target lesions measured under RECIST 1.1. Bars are sorted from the largest increase on the left to the largest decrease on the right, so tumor growth sits above the zero line and shrinkage hangs below it, and the falling profile gives the plot its name. Colors usually show each patient’s best overall response, and dashed lines at −30% and +20% mark the RECIST thresholds for partial response and progression. A waterfall plot shows depth of response for every patient at once, which a single response rate hides.

How to read a waterfall plot

• Each bar is one patient. Its height is the best (most negative) percentage change from baseline seen at any assessment, not the change at a fixed time.
• Bars below zero are patients whose target lesions shrank. A bar at −100% means every target lesion disappeared.
• The −30% line marks the size threshold for a RECIST partial response. The share of bars below it is a quick visual check, but the response rate must come from the confirmed best overall response, not from counting bars.
• The +20% line is a visual guide. RECIST progression is measured from the smallest sum on study (the nadir), not from baseline, and also needs a 5 mm absolute increase, so a bar can stay below +20% from baseline while the patient progressed.
• Colors carry information the bar height cannot: a bar below −30% can be stable disease (an unconfirmed response), and a bar below zero can be progressive disease (a new lesion or non-target progression).
• A waterfall plot shows depth of response, not durability. Pair it with a Kaplan-Meier curve of progression-free survival or a swimmer plot of time on treatment.

RECIST 1.1 thresholds behind the reference lines

• Complete response (CR): all target lesions have disappeared, and any pathological lymph node has shrunk to under 10 mm in short axis. Because nodes are still measured, a CR bar can sit slightly above −100%.
• Partial response (PR): at least a 30% decrease in the sum of target lesion diameters, taking the baseline sum as reference.
• Progressive disease (PD): at least a 20% increase in the sum of diameters, taking the smallest sum on study as reference, with an absolute increase of at least 5 mm. One or more new lesions also count as progression.
• Stable disease (SD): neither enough shrinkage for PR nor enough growth for PD, taking the smallest sum on study as reference.
• In single-arm trials where response is the primary endpoint, CR and PR must be confirmed by a repeat assessment, normally at least 4 weeks later. An unconfirmed response is reported as SD.
• Immunotherapy trials may use iRECIST, which adds unconfirmed and confirmed progression categories (iUPD, iCPD) because tumors can grow before they shrink.

What does this waterfall plot generator calculate?

The generator reads your best percentage change column, rejects values below −100% (a tumor cannot shrink by more than all of its size), and sorts the bars. If you map a best overall response column, it normalizes CR, PR, SD, PD and NE in any spelling, counts each category, and reports the objective response (CR + PR) and disease control (CR + PR + SD) counts with the denominator printed, using only the responses in your file. It flags any PR or CR bar above −30%, which RECIST cannot produce, and notes PD bars with shrinkage and SD bars below −30%, which RECIST allows. Without a response column, bars are colored by change band and the figure says the bands are not RECIST responses. Bars can be capped (for example at +100%) with the true value printed above them. The image model then draws these exact bars; it does not run statistical tests or add p-values.

Waterfall chart for budgets, energy balances, and other running totals

The same name is used for a different chart: the cumulative, or bridge, waterfall chart from finance. It starts from an opening value, adds increases and subtracts decreases as floating bars, and ends at a total, so each bar starts where the previous one ended. Add a step type column with start, change, subtotal and total values and the generator switches to this mode: it computes the running total, places each floating bar, recomputes every subtotal and total, and warns you when a total in your file does not match the sum of its steps. In research it suits a grant budget, an energy or mass balance, a carbon budget, or a decomposition of a change into its parts.

What data does a waterfall plot need?

• Tumor response: one row per patient with an ID and a best percentage change column (values such as −42, 18 or −100; a % sign is fine). Optional: best overall response (CR, PR, SD, PD, NE or the full words) and a cohort, arm, dose level or biomarker column.
• Compute best percentage change as (smallest post-baseline sum − baseline sum) / baseline sum × 100 for each patient before uploading. CDISC ADaM datasets usually hold it as PCHG in an ADTR-style dataset.
• Patients without a post-baseline scan have no bar. Leave them out of the file and report how many there were in the caption.
• Preclinical studies work the same way: percentage change in tumor volume per animal, with the treatment group in the response column so bars are colored by group.
• Cumulative waterfall chart: one row per step with a label, an amount, and a type (start, change, subtotal, total). Leave the amount empty on subtotal and total rows; the generator fills them in.

How to report a waterfall plot in a paper

• State the measure on the y-axis in full: best percentage change from baseline in the sum of target lesion diameters.
• Give the number of patients plotted and the number left out for having no post-baseline assessment.
• Say who assessed response (investigator or blinded independent central review), which criteria (RECIST 1.1, iRECIST, mRECIST), and whether responses were confirmed.
• Explain colors that do not match bar height, for example a footnote marking PD due to new lesions.
• Keep the y-axis honest: if you cap bars, mark the capped bars and give their true values.