This page provides an independent statistical analysis and educational interpretation of publicly reported results. ClinicalTrials.gov provides the official trial registry record.
1. Trial at a Glance
ARTEMIS is a randomized, parallel, open-label phase 3 treatment trial evaluating bevacizumab in combination with erlotinib versus erlotinib alone in patients with NSCLC harbouring activating EGFR mutations. The registry lists progression-free survival by independent review committee assessment as the primary endpoint.
| Feature | ARTEMIS |
|---|---|
| Trial acronym | ARTEMIS |
| ClinicalTrials.gov identifier | NCT02759614 |
| Phase | Phase 3 |
| Status | Completed |
| Condition | NSCLC |
| Enrollment | 311 |
| Allocation | Randomized |
| Design model | Parallel |
| Masking | None |
| Primary purpose | Treatment |
| Lead sponsor | Guangdong Association of Clinical Trials |
2. Clinical Question
The central statistical question is whether adding bevacizumab to erlotinib improves progression-free survival compared with erlotinib alone in patients with NSCLC harbouring activating EGFR mutations.
Population
Patients with NSCLC harbouring activating EGFR mutations.
Intervention
Bevacizumab in combination with erlotinib.
Comparator
Erlotinib alone.
Primary question
Does bevacizumab combined with erlotinib improve independent review committee assessed progression-free survival compared with erlotinib alone?
3. Trial Design
ARTEMIS uses a randomized, parallel-group phase 3 design with two treatment arms. The registry identifies the primary purpose as treatment and the masking status as none, making this an open-label randomized comparison.
Bevacizumab + Erlotinib
- Bevacizumab
- Erlotinib
- Combination treatment
Erlotinib Alone
- Erlotinib
- Single-agent comparator
The randomized parallel structure is important statistically because the treatment groups are defined by assignment at the beginning of the comparison. The principal efficacy question is therefore a between-group time-to-event comparison rather than a within-patient comparison.
4. Trial Timeline
Trial start
The registry lists April 1, 2016 as the study start date.
Primary completion
The registry lists June 1, 2018 as the primary completion date.
Study status
The registry lists ARTEMIS as completed.
5. Endpoints
| Endpoint | Registry definition | Time frame |
|---|---|---|
| Progression-free survival (PFS) by IRC | To evaluate the efficacy of bevacizumab combined with erlotinib compared with erlotinib alone in patients with NSCLC harbouring activating EGFR mutations, as measured by independent review committee assessed progression-free survival using RECIST v1.1. | The primary PFS analysis will be performed when approximately 224 PFS events (disease progression or death, whichever occurs first) have occurred, which is estimated to occur at approximately 18 months after enrollment of the last patient. |
What the endpoint measures
Progression-free survival is a time-to-event endpoint. For this trial, the registry defines the event as either disease progression or death, whichever occurs first. A participant who has neither event by the relevant follow-up point contributes information through the time at which follow-up is censored under the prespecified analysis rules.
The use of independent review committee assessment is also statistically important. The primary endpoint is not simply a treatment investigator's assessment; it is explicitly defined around an independent review process using RECIST v1.1.
6. Statistical Methodology
Event-driven analysis
The primary PFS analysis is event-driven rather than based solely on a fixed calendar date. The registry specifies approximately 224 PFS events as the trigger for the primary analysis and estimates that this will occur at approximately 18 months after enrollment of the last patient.
A PFS event is defined in the registry as disease progression or death, whichever occurs first.
An event-driven design is natural for a time-to-event endpoint because the precision of a treatment-effect estimate depends strongly on the number of observed events. Calendar time alone does not determine how much statistical information has accumulated.
Independent review committee assessment
The registry identifies progression-free survival by independent review committee assessment as the primary endpoint and specifies RECIST v1.1 for the assessment. This approach is intended to provide a standardized efficacy assessment that is separated from routine treatment decisions made by the treating investigators.
Typical time-to-event analysis
For a randomized trial with a PFS endpoint, a conventional analysis would estimate the distribution of PFS using the Kaplan-Meier method and compare randomized treatment groups using a time-to-event hypothesis test such as the log-rank test. A Cox proportional-hazards model is commonly used to estimate a hazard ratio and its confidence interval.
Those methods provide complementary information. Kaplan-Meier estimation describes the event-free probability over time, the log-rank framework evaluates the treatment-group separation across follow-up, and the Cox model summarizes the relative event rate through a hazard ratio.
A hazard ratio is a relative time-to-event measure. It is not itself a percentage of patients cured, a relative reduction in median PFS, or an absolute difference in the probability of remaining progression-free at a particular time.
7. Planned Analysis
The registry specifies progression-free survival by independent review committee as the primary endpoint. The endpoint is measured from the time-to-event perspective, with disease progression or death, whichever occurs first, defining the event. The primary PFS analysis is planned when approximately 224 such events have occurred, estimated at approximately 18 months after enrollment of the last patient.
Primary statistical information available from the registry
Approximately 224 disease-progression-or-death events are specified as the trigger for the primary PFS analysis.
For an endpoint of this type, the usual statistical framework would use time-to-event methods that account for censoring. Kaplan-Meier estimation can describe PFS over time, while a log-rank comparison and Cox proportional-hazards model can be used to compare randomized groups and estimate a relative treatment effect. The precise formal testing procedure, model specifications, multiplicity strategy, and confidence-interval construction are not reported in the registry information summarized here.
8. Statistical Methods Explained
Why is progression-free survival a time-to-event endpoint?
PFS incorporates both whether an event occurs and when it occurs. Two patients can both eventually experience progression, but the timing of progression contains additional information. Treating PFS simply as a binary outcome would discard that timing information.
Why are both progression and death counted as events?
The registry defines the PFS event as disease progression or death, whichever occurs first. This prevents death before documented progression from being treated as if the patient remained progression-free indefinitely.
Why use an independent review committee?
An independent review committee provides an assessment of tumor progression that is separate from the treating investigator's routine clinical assessment. In a trial where progression is the primary endpoint, consistency of radiologic assessment is particularly important because the event definition directly determines the time-to-event outcome.
Why is the analysis event-driven?
The registry specifies approximately 224 PFS events rather than simply specifying a single fixed follow-up date for the primary analysis. For survival endpoints, the number of observed events is a major determinant of the information available for estimating a treatment effect.
What does a hazard ratio represent?
A hazard ratio compares the estimated instantaneous event rates between treatment groups within a time-to-event model. If a hazard ratio were below 1 for bevacizumab plus erlotinib versus erlotinib alone, it would indicate a lower estimated instantaneous rate of progression or death under that model. It would not mean that the same percentage of every patient experienced a corresponding reduction in risk.
Why does censoring matter?
Some participants may reach the end of observable follow-up without experiencing progression or death. Survival methods allow these participants to contribute information up to their censoring time rather than assuming that they subsequently experienced no event or that they experienced an event at a particular unobserved time.
9. Interpreting the Primary Endpoint
Relative effect
A hazard ratio, if reported, would summarize the relative instantaneous rate of progression or death between the two randomized treatment groups.
Absolute effect
Kaplan-Meier estimates at specified times would show the absolute probability of remaining free from progression or death at those times.
Event timing
The endpoint incorporates when progression or death occurs, not merely whether an event eventually occurred.
Precision
A confidence interval around a treatment-effect estimate would describe statistical uncertainty and would be essential for understanding the precision of the estimated effect.
These measures should be interpreted together. A hazard ratio provides a compact relative summary, while the Kaplan-Meier curve and time-specific survival estimates provide information about the absolute course of PFS. Neither measure alone describes every aspect of treatment effect.
10. Hazard Ratios: What They Do — and Do Not — Mean
For ARTEMIS, the relevant hypothetical treatment comparison is bevacizumab plus erlotinib versus erlotinib alone. A hazard ratio below 1 would indicate a lower estimated instantaneous rate of disease progression or death in the combination group under the fitted time-to-event model.
A hazard ratio is not the same as a relative risk, an absolute risk difference, a difference in median PFS, or the percentage of patients who benefit. For example, a hazard ratio of 0.70 would not mean that 30% of patients were cured or that median PFS was increased by exactly 30%.
A confidence interval quantifies uncertainty around an estimated treatment effect. A narrow interval indicates greater statistical precision than a wide interval, all else being equal. It does not describe the range of outcomes that individual patients will experience.
A p-value addresses evidence against a specified null hypothesis under the statistical testing framework. It does not describe the magnitude of a treatment effect. Effect size and statistical uncertainty should be reported directly through measures such as the hazard ratio and its confidence interval.
11. Censoring and the Kaplan-Meier Framework
Progression-free survival data are commonly subject to right censoring. A participant may leave the observable risk set without having experienced the defined event, or may reach a prespecified assessment boundary while remaining progression-free. The Kaplan-Meier method is designed to use the information available before censoring.
Here, di is the number of events at event time ti, and ni is the number of participants at risk immediately before that time.
The key point is that a censored participant is not automatically classified as progression-free for all future time. Instead, that participant contributes information through the censoring time according to the prespecified analysis rules.
12. RECIST v1.1 and Endpoint Assessment
The registry specifies RECIST v1.1 as the framework for the independent review committee assessment of progression-free survival. This matters because a time-to-progression component of PFS depends on a consistent definition of what constitutes disease progression.
| Component | Role in ARTEMIS |
|---|---|
| Progression-free survival | Primary endpoint |
| Independent review committee | Assesses the primary PFS endpoint |
| RECIST v1.1 | Specified assessment framework |
| Death | Counts as a PFS event when it occurs before documented disease progression |
The endpoint therefore combines radiologic disease assessment with survival status. This is one reason PFS analysis requires a defined event algorithm rather than simply counting radiographic responses.
13. Randomization and Causal Interpretation
Randomization is the central design feature supporting a causal comparison between bevacizumab plus erlotinib and erlotinib alone. Because treatment assignment is randomized, baseline prognostic factors are expected to be balanced in expectation rather than selected according to clinical characteristics.
The absence of masking does not eliminate the value of randomization, but it is relevant to how the trial is interpreted. Treatment assignment is known to participants and investigators. For the primary endpoint, the use of independent review committee assessment provides an additional layer of separation between treatment assignment and progression assessment.
Randomization
Creates the principal basis for comparing outcomes between the two treatment groups without relying on observational adjustment alone.
No masking
The registry identifies masking as none, so treatment assignment is not blinded.
Independent assessment
The primary endpoint uses independent review committee assessment, which is particularly relevant for an imaging-based progression endpoint.
Parallel design
Participants remain in their randomized treatment-group comparison rather than serving as their own controls in a crossover design.
14. What the Registry Does and Does Not Establish
The registry establishes the principal design and endpoint framework: ARTEMIS is a completed randomized phase 3 parallel trial with 311 participants, two treatment arms, no masking, and a primary PFS endpoint assessed by an independent review committee using RECIST v1.1.
It also specifies the event-driven trigger for the primary analysis: approximately 224 PFS events, defined as disease progression or death, whichever occurs first, with an estimated timing of approximately 18 months after enrollment of the last patient.
The registry record summarized here does not report a numerical PFS treatment effect, confidence interval, or p-value. Those quantities would normally be central to the statistical interpretation of a completed randomized efficacy analysis.
15. Limitations and Interpretation Issues
- No posted formal efficacy analysis: the registry record identifies the primary endpoint and its planned event-driven analysis but does not provide an observed treatment-effect estimate, confidence interval, or p-value.
- Open-label design: masking is listed as none. Knowledge of treatment assignment can affect aspects of clinical management and participant experience, although the primary endpoint uses independent review committee assessment.
- Time-to-event assumptions: a Cox proportional-hazards analysis, if used, relies on a proportional-hazards framework for its standard single-hazard-ratio interpretation.
- Censoring: PFS analyses depend on prespecified censoring and assessment rules. Those detailed rules are not reported in the registry information summarized here.
- Event definition: PFS combines disease progression and death. The resulting endpoint therefore represents a composite time-to-event outcome rather than progression alone.
- Analysis specification: the registry information does not specify the complete statistical analysis plan, including the exact hypothesis-testing procedure, model covariates, confidence-interval construction, or multiplicity strategy.
- Clinical interpretation: an eventual statistical difference in PFS would not by itself describe overall survival, quality of life, toxicity, or every dimension of clinical benefit.
16. Why This Trial Matters Statistically
ARTEMIS is a useful teaching case because its registry design illustrates the structure of an event-driven randomized time-to-event trial without conflating trial design with observed efficacy results.
| Concept | How it appears in ARTEMIS |
|---|---|
| Randomization | Participants are assigned randomly to two treatment arms. |
| Parallel-group design | The registry identifies the design model as parallel. |
| Phase 3 testing | The study is listed as phase 3. |
| Time-to-event endpoint | Progression-free survival is the primary endpoint. |
| Composite event | PFS counts disease progression or death, whichever occurs first. |
| Independent assessment | The primary endpoint is assessed by an independent review committee. |
| RECIST v1.1 | The registry specifies RECIST v1.1 for the primary PFS assessment. |
| Event-driven analysis | The primary analysis is planned after approximately 224 PFS events. |
| Censoring | Time-to-event analysis requires handling participants who have not experienced the event by their last assessment. |
| Hazard ratio | A conventional time-to-event analysis can summarize the relative rate of progression or death between randomized groups. |
17. Clinical Interpretation vs Statistical Interpretation
Statistical interpretation
The prespecified primary endpoint is an event-driven time-to-event measure. Its analysis is intended to compare the distribution of progression-free survival between randomized treatment groups after approximately 224 PFS events have accumulated.
Clinical interpretation
The clinical question concerns whether adding bevacizumab to erlotinib changes the course of disease progression or death in patients with NSCLC harbouring activating EGFR mutations. PFS addresses that question over time but does not encompass every clinical outcome.
Keeping these interpretations separate is important. The statistical analysis determines how evidence for a treatment difference is quantified; the clinical interpretation determines what the endpoint represents in the context of the disease and treatment strategy.
18. Primary Endpoint: A Step-by-Step Statistical View
This sequence illustrates why the analysis is not simply a comparison of the number of patients who eventually progressed. The timing of progression or death contributes directly to the statistical information in the endpoint.
19. Interpreting a Future PFS Result
If a formal PFS analysis reports a hazard ratio, the estimate should be read together with its confidence interval and the prespecified testing framework. A hazard ratio below 1 would indicate a lower estimated instantaneous rate of progression or death for bevacizumab plus erlotinib relative to erlotinib alone, while a hazard ratio above 1 would indicate a higher estimated instantaneous rate.
| Reported quantity | What it tells the reader |
|---|---|
| Hazard ratio | Relative treatment effect on the instantaneous event rate under the fitted time-to-event model. |
| 95% confidence interval | Statistical uncertainty around the estimated treatment effect. |
| P-value | Evidence against a specified null hypothesis under the prespecified testing framework. |
| Median PFS | A time-point summary of when the estimated progression-free probability reaches 50%, when that point is estimable. |
| Kaplan-Meier estimates | Estimated probability of remaining free from progression or death at specified times. |
No one of these measures is sufficient on its own. The most informative statistical presentation combines a relative effect measure with its uncertainty and absolute time-to-event estimates.
20. Record Summary
ARTEMIS is a completed randomized phase 3 parallel-group trial enrolling 311 participants to compare bevacizumab in combination with erlotinib against erlotinib alone in NSCLC patients harbouring activating EGFR mutations. Its primary endpoint is progression-free survival by independent review committee, assessed using RECIST v1.1, with disease progression or death, whichever occurs first, defining the PFS event.
The registry specifies an event-driven primary analysis after approximately 224 PFS events, estimated to occur at approximately 18 months after enrollment of the last patient. This makes ARTEMIS a useful example of how randomized clinical trials can be designed around the accumulation of time-to-event information rather than a fixed calendar endpoint.
For statistical interpretation, the central concepts are randomization, independent endpoint assessment, RECIST-based progression, censoring, Kaplan-Meier estimation, comparative time-to-event testing, and hazard-ratio interpretation. The registry record summarized here does not contain a posted formal statistical analysis of the primary endpoint, so the available record supports interpretation of the planned analysis rather than a numerical conclusion about treatment effect.
21. Sources
- ClinicalTrials.gov: ARTEMIS, NCT02759614.