I-SPY 2.2 And Response-Based Treatment Paths

Why Treatment Paths Matter

In neoadjuvant breast cancer treatment, therapy is given before surgery, allowing clinicians to observe tumor response while treatment is under way. I-SPY 2.2 uses that opportunity to organize care in sequential treatment blocks: patients begin with a treatment, undergo early response assessment, and may then proceed to surgery or receive further treatment. The first block can include an investigational regimen; later blocks include additional therapy according to the protocol.

The clinical and statistical question is therefore not only whether a drug has activity. It is also whether a response-guided treatment path—a rule connecting initial treatment and early response to subsequent care—produces useful outcomes. A patient who responds early may receive fewer treatment blocks than a patient whose tumor does not respond sufficiently. Those paths differ in treatment exposure as well as in the patients who follow them.

I-SPY 2.2 uses serial breast MRI, including functional tumor volume (FTV), as an early measure of tumor response. The protocol’s response algorithm uses such information to guide decisions about subsequent blocks. The precise decision rules and treatment details are protocol-specific; an early MRI result is not itself the final clinical outcome. Pathologic complete response (pCR), assessed at surgery under the trial’s definition, is an important outcome for evaluating whether a treatment strategy is promising.

From an Early Measure to a Treatment Decision

An early response measure is useful because it can inform treatment before the full course is complete. For example, a change in FTV can summarize how the MRI-measured functional tumor volume has changed from baseline. If \(FTV_0\) is the baseline measurement and \(FTV_1\) is a later measurement, the percentage reduction is

A positive value indicates a reduction; zero indicates no measured change; and a negative value indicates an increase. This calculation describes a change in the imaging measure, not a percentage probability of pCR. In I-SPY 2.2, the treatment decision uses a prespecified response algorithm rather than an informal judgment based on one number. The algorithm relates early response information to the possibility of later outcomes and determines whether further treatment is indicated.

A useful early measure must be collected at the right time, measured consistently, and interpreted in the context of the treatment rule. If MRI scans are delayed or unavailable, the decision may be made with incomplete information. Measurement variation can also make an apparent change differ from the underlying tumor change. In addition, patients assessed at different times may not have directly comparable FTV changes. These are reasons to document assessment timing and data completeness and to apply the protocol’s rules consistently.

Why the Path Is the Unit of Interpretation

Suppose a patient receives an investigational regimen in an initial block and then receives more treatment because the early response is insufficient. The final outcome reflects the initial regimen, the additional treatment, the timing and quality of response measurement, and the patient’s characteristics. It cannot generally be attributed to the initial drug alone. Conversely, a patient who proceeds to surgery after an early response has had a different treatment exposure. Comparing these patients as though they had received interchangeable versions of one therapy would obscure the decision rule that created the difference.

This distinction matters because response is measured after treatment begins. Patients who proceed to a later block or to surgery are selected partly by their early response. Early response is also related to prognosis. Thus, a simple comparison of patients who did and did not receive a later block mixes treatment effects with differences in the patients selected for that path.

Defining an Estimand for a Response-Guided Policy

A causal question needs a clearly stated comparison. Let \(g\) denote a fully specified policy, including the initial treatment and the rules for subsequent treatment. Let \(Y^g\) denote the outcome a patient would have under policy \(g\), such as an indicator equal to 1 for pCR and 0 otherwise. The policy-level pCR probability is \(P(Y^g=1)\). Comparing policies \(g_1\) and \(g_0\) can be expressed as a difference in pCR probabilities:

This contrast concerns the policies as complete strategies. If the policies differ in their initial treatments, their response rules, or both, the contrast includes those differences. It is not automatically an estimate of the effect of a single drug independent of subsequent care.

A simple justification shows why baseline randomization is valuable. If patients are randomly assigned to one of two prespecified policies, assignment is independent of their potential outcomes. With consistent implementation and complete outcome ascertainment, the observed pCR proportion in each assigned group estimates the pCR probability under that group’s policy. The difference between those observed proportions then estimates the policy contrast, subject to sampling uncertainty. By contrast, grouping patients by the path they actually followed does not preserve that randomization: the observed path is determined in part by post-assignment response.

Interpreting the Design and Its Limits

The response-guided structure can make treatment more adaptive: early information can direct patients toward different subsequent care rather than requiring every patient to receive an identical sequence. It also changes what a trial result means. Evidence that a treatment path has promising outcomes does not, by itself, prove that every component is beneficial, that a different response rule would work as well, or that the strategy improves long-term outcomes. pCR is an important neoadjuvant endpoint, but it does not answer every question about recurrence, survival, toxicity, or quality of life.

A sound analysis should therefore make the target explicit: Is the question about a complete response-guided policy, an initial regimen under a common policy, or the effect of a particular later treatment decision? The answer determines which comparison is appropriate. Results should also be interpreted in light of the population and biomarker-defined groups for which the response model and treatment rules were developed. A rule that performs well in one population need not transfer unchanged to another.

A common mistake is to compare outcomes among patients who reached different blocks and call the difference a treatment effect. The reasoning is tempting: one group received more treatment, so a difference seems to measure the effect of that treatment. The correction is that early response helps determine who receives the next block and also predicts outcome. The groups are therefore not exchangeable by design. Report those comparisons as descriptions of observed paths, or use a causal method only when its assumptions and the required data are justified.