Why Response Endpoints Need Careful Definition
In oncology, response endpoints can summarize whether a tumor shrank or disappeared after treatment. In a clinical trial, scheduled imaging and prespecified assessment rules can make response classification relatively consistent across participants. Routine-care records are different: scans may occur at variable intervals, reports may not include lesion-level measurements, and a clinician's note may use terms such as “improving” without defining a formal response category. The resulting real-world endpoint can be useful, but it is not automatically equivalent to a trial endpoint.
This distinction matters when real-world outcomes are considered alongside trial results or used in an external-control analysis. Differences in assessment rules can create apparent differences in response even when underlying tumor changes are similar. Covariate adjustment does not, by itself, make two differently defined response outcomes comparable.
Trial Response Criteria as a Reference Point
RECIST 1.1 is a standard framework for classifying response in solid tumors. It uses prespecified imaging assessments, baseline measurements of selected target lesions, and rules for target lesions, non-target lesions, and new lesions. For target lesions, the sum of longest diameters is commonly used; lymph nodes are measured by short axis. The same measurement approach and lesion set must be applied consistently for the thresholds to be meaningful.
For a measurable target-lesion sum \(S_0\) at baseline and sum \(S_t\) at a later assessment, the percentage change from baseline is \(100(S_t-S_0)/S_0\). A partial response (PR) requires at least a 30% decrease in the sum from baseline. Progressive disease (PD) for target lesions requires at least a 20% increase from the smallest sum recorded since treatment began (the nadir), with an absolute increase of at least 5 mm. Complete response (CR) requires disappearance of target lesions, with lymph nodes reduced below the specified size threshold; the overall category also depends on non-target disease and whether new lesions have appeared. Stable disease (SD) is assigned when neither response nor progression criteria are met, subject to the assessment rules.
Best overall response (BOR) is the best response category recorded across the defined assessment period, subject to any applicable confirmation rules. Overall response rate (ORR) is the proportion of the defined analysis population whose BOR is CR or PR. In a trial, the protocol specifies the assessment schedule, analysis population, and rules for unconfirmed responses. RECIST response is therefore not just a percentage change: it is a classification under a complete assessment framework.
Routine records may support a RECIST-like calculation if they preserve the needed lesion measurements and assessment details. More often, real-world response is based on clinician-assessed categories or extracted statements in records. Such outcomes should be named and described distinctly—for example, clinician-documented real-world response—rather than presented as RECIST response when the RECIST requirements were not met.
Worked Example: Checking Whether Measurements Meet the PR Threshold
In a hypothetical record, a patient has a baseline target-lesion sum of 80 mm. At a later scan the recorded sum is 59 mm. Assuming the same target lesions and measurement method were used, determine whether the target-lesion change meets the RECIST PR size threshold.
The percentage change is
The sum decreased by 26.25%, which is less than the 30% decrease required for PR. The target-lesion measurements alone therefore do not meet the PR threshold. This does not establish the patient's overall response category: non-target disease, new lesions, and other RECIST rules must also be considered. A note describing the tumor as “markedly smaller” would not substitute for the threshold calculation.
Constructing an Endpoint From Routine Records
Before extracting response, specify what the endpoint is intended to represent. A practical protocol should define the eligible treatment episode and baseline window, the post-treatment assessment period, the evidence that qualifies as response, and how assessments are ordered to determine BOR. It should also state whether a response must be documented at more than one assessment. Confirmation may be required by a trial protocol, particularly in some nonrandomized settings, but it should not be assumed to be present—or imposed retrospectively—without a stated rule.
The following distinctions are especially important:
- Measurement-based versus clinician-assessed: A measured lesion sum may permit application of specified size thresholds. A clinician's documented “response” can support a clinician-assessed endpoint, but does not show that all measurement-based criteria were satisfied.
- Observed response versus BOR: A single scan can show a response at that assessment. BOR requires reviewing the eligible sequence of assessments and applying the stated rules, including any confirmation requirement.
- Response status versus timing: The scan date records when response was observed, not necessarily when the biological change began. If scans are months apart, the onset of response is not known to that precision.
- Documented nonresponse versus no assessment: A record with no post-baseline scan does not establish that the patient had no response. Coding it as nonresponse is an analysis convention that must be disclosed, not a clinical finding.
Scan timing affects the chance to observe a response. A patient assessed frequently has more opportunities for a response to be recorded than a patient with only one late scan. Conversely, a response that appears and resolves between widely spaced scans may never be documented. If treatment is stopped, the patient becomes too unwell for imaging, or scans are ordered in response to symptoms, assessment timing may also depend on clinical course. These mechanisms can make the observed endpoint selective rather than a neutral snapshot of response.
Worked Example: Interpreting Response Across Uneven Scan Timing
In a hypothetical cohort, a patient has a baseline scan on January 5. A February 20 routine scan is documented as a partial response by the treating oncologist. A further scan is not obtained until May 30, when progressive disease is recorded. Suppose the endpoint rule defines response as any documented CR or PR during the first six months, without requiring confirmation.
The patient meets that operational response rule because a PR was documented within six months. The later PD does not erase a previously observed response when the endpoint is “any response during the window” or BOR under a rule that includes the February assessment. However, if the endpoint requires confirmed response, the February PR cannot be called confirmed from the information given: the record provides no qualifying confirmatory assessment. The long interval before the next scan also means the date and course of progression are not precisely observed.
If the intended endpoint instead requires response to be present at a fixed six-month assessment, the record does not establish that response: the May 30 scan documents PD. These classifications differ because the endpoint definitions differ, not because the underlying records changed.
Denominators, Missing Assessments, and Uncertainty
An ORR calculation requires both a response rule and a denominator. A trial-style all-treated denominator answers a different question from an evaluable-patient denominator. The latter can overstate response if patients without assessments are excluded and their missingness is related to health, treatment discontinuation, or access to imaging. A transparent report should give the numerator and denominator and describe how unevaluable patients were handled.
For a simple binary endpoint, let \(R\) be the number classified as responders and \(N\) the analysis denominator. The observed response proportion is \(R/N\). This arithmetic does not solve the missing-data problem: choosing \(N\) and assigning or excluding missing outcomes are substantive endpoint decisions. In a Bayesian analysis, a prior can represent uncertainty about a response probability, but it cannot recover undocumented response status without assumptions about the missing assessments.
Worked Example: How Missing Assessments Change the Reported Rate
A hypothetical group has 40 treated patients. Records document response for 12 patients, documented nonresponse for 19, and no evaluable post-baseline assessment for 9. Calculate rates under two explicit conventions and show the range implied by the unknown outcomes.
Among the 31 patients with an evaluable assessment, the response rate is
or 38.7%. This is an evaluable-patient rate and describes only those with an assessment. If all 40 treated patients are used as the denominator and the 9 with no assessment are counted as nonresponders, the rate is
or 30.0%. That value depends on the stated nonresponder convention; it does not mean the nine patients were clinically shown not to respond. If all nine unknown outcomes were responses, the highest rate consistent with these records would be
or 52.5%. Thus, under the all-treated denominator, the records alone permit a range from 30.0% to 52.5% if the missing outcomes are left unrestricted. The range is not a confidence or credible interval; it is a simple bound showing the impact of unknown classifications.
Comparison With Common Alternatives
| Approach | What it uses | Main strength and limitation |
|---|---|---|
| RECIST-based trial response | Protocol-scheduled imaging, lesion measurements, and formal rules | Standardized classification; may not be reproducible from routine records lacking measurements or comparable scan timing |
| Clinician-assessed real-world response | Response categories or judgments documented in routine care | Reflects ordinary clinical documentation; terminology and completeness can vary across clinicians and records |
| Radiology-report extraction | Impression statements or measurements in scan reports | Uses imaging documentation directly; reports may not contain all lesions, formal thresholds, or a consistent assessment framework |
| Treatment-change proxy | Whether therapy was continued, changed, or stopped | May be readily available, but treatment decisions also reflect toxicity, preference, access, and other clinical considerations; it is not itself a response endpoint |
These approaches are alternatives for different data conditions, not interchangeable labels for the same measurement. If a real-world response estimate is compared with trial ORR, the comparison should state the differences in assessment schedule, response source, confirmation, population denominator, and missing-data rule. A numerical similarity does not establish that the endpoints were measured equivalently.
Diagnostics and Common Pitfalls
Several checks help determine whether routine records support the intended response endpoint. Review whether baseline imaging precedes treatment and falls within the specified window; whether measurements or response terms are actually present; how often assessments occur; and how many patients have no post-baseline assessment. Tabulate assessment counts and timing, including the proportion assessed within the endpoint window. Inspect a sample of records to verify that extraction rules distinguish explicit response statements from vague descriptions and treatment decisions.
- Calling a clinician note “RECIST response.” A note can support clinician-assessed response, but RECIST classification requires its specified lesion, measurement, and overall-response rules.
- Assuming no recorded response means nonresponse. Missing documentation is not evidence of no biological response. State and justify the handling rule, then examine how conclusions change under plausible alternatives.
- Comparing raw ORRs as if scans were scheduled alike. Different assessment frequency and follow-up windows can change the opportunity to record response. Describe these differences before interpreting a contrast.
- Using treatment continuation as proof of response. Treatment may continue despite lack of tumor shrinkage or stop despite response. Treat therapy decisions as context, not a substitute response measure.
In external-control work, response endpoints are especially vulnerable to differences in documentation and assessment opportunity. Adjustment for measured patient characteristics addresses neither an undefined endpoint nor missing lesion measurements. If a response definition cannot be applied consistently, report what each source actually measures and qualify any comparison rather than implying a common endpoint.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A target-lesion sum decreases from 100 mm at baseline to 68 mm at follow-up. What is the percentage change, and does it meet the target-lesion PR size threshold?
- In a hypothetical treated group of 50 patients, 15 have documented response, 25 have documented nonresponse, and 10 have no assessment. Calculate the evaluable-patient response rate and the all-treated rate if missing assessments are counted as nonresponse.
- A record contains one clinician-documented PR and no later scan. What can be concluded if the endpoint requires any documented response? What can be concluded if it requires confirmed response?
- Why can a routine-care ORR differ from a trial ORR even when the patient groups have similar tumor biology?
Solutions. (1) \(100(68-100)/100=-32\%\), a 32% decrease, so the target-lesion size threshold for PR is met; overall classification still requires the other relevant criteria. (2) The evaluable rate is \(15/(15+25)=15/40=37.5\%\). The all-treated rate under the stated missing-as-nonresponse convention is \(15/50=30.0\%\). (3) The patient satisfies an endpoint defined as any documented response, but the record does not establish confirmation because no later qualifying assessment is available. (4) Scan frequency, assessment windows, response documentation, confirmation rules, denominator choice, and missing assessments can all differ and alter the observed rate.