Background: MRI measured functional tumor volume (FTV) canpredict pathologic complete response (pCR) to neoadjuvant chemotherapy (NAC) asearly as 3 weeks after treatment initiation (1), indicating the potential ofusing MRI to guide treatment de-escalation in clinical trials. We developed MRIbased, subtype-specific, predictive models for informing a de-escalationstrategy in I-SPY 2 after 12 weeks of NAC.
Methods: I-SPY 2 patients underwent MRI exams atpre-treatment, early treatment (3 wks), inter-regimen (12 wks), andpre-surgery. pCR was assessed at surgery for each patient. FTV was calculatedsemi-automatically in each MRI(2). The MRI model for predicting pCR was builtat inter-regimen, by selecting FTV predictors (pre-treatment FTV and percentagechanges at early treatment and inter-regimen compared to pre-treatment)separately for each cancer subtype as defined by hormone receptor (HR) andhuman epidermal growth factors receptor 2 (HER2) status. The subtype-specificmodel was finalized by achieving the best predictive performance evaluated byarea under the receiver operating characteristic curve (AUC) for predictingpCR. A poatient's predicted probability for pCR above a specific threshold wasconsidered a positive test.The therapy de-escalation strategy focuses onfinding positives (pCRs) while minimizing false positives (type I error –firstpriority) and false negatives (type II error –second priority), i.e. maximizingsensitivity while controlling for positive predictive value (PPV: proportion ofpatients with test positive who achieved pCR).However, an increased probabilitythreshold will decrease the number of patients with positive tests, possibly increasingPPV but adversely affecting sensitivity. This study shows the tradeoff betweenPPV and sensitivity when the MRI prediction model is applied to therapyde-escalation.
Results: 814 patients enrolled in I-SPY 2 between May 2010and November 2016 were included in the analysis. Median age was 49 (range: 24–77) years. The pCR rate was 36% (289/814). Table 1 shows patient number andpCR rate by HR/HER2 subtype. The subtype-specific MRI models consist of thepredictors: change of FTV (∆FTV) at inter-regimen for HR+/HER2-and HR-/HER2+;∆FTV at early treatment for HR+/HER2+; pre-treatment FTV and ∆FTV atinter-regimen for triple negatives. The highest probability varied by subtype:0.24 for HR+/HER2-, 0.61 for HR+/HER2+, 0.73 for HR-/HER2+, 0.68 for triplenegatives. The maximum PPV was 67% for HR+/HER2-and 100% for all other subtypes. Table 1 shows the tradeoffbetween PPV and sensitivity in each subtype when the probability threshold was chosenat the 1st, 2nd(median), and 3rdquartile.
Conclusions: Our data demonstrate that PPV and sensitivityvary by breast cancer subtype when the probability threshold generated by MRImodel increases from low to high quartile. Results from this study suggest thatthe probability threshold for recommending treatment de-escalation should beselected carefully based on breast cancer subtype. Imaging results will becombined with core biopsy information obtained at the 12-week timepoint tofurther improve overall accuracy.
1. Hylton NM, Blume JD, Bernreuter WK, Pisano ED, Rosen MA,Morris EA, et al. Locally advanced breast cancer: MR imaging for prediction ofresponse to neoadjuvant chemotherapy--results from ACRIN 6657/I-SPY TRIAL.Radiology. 2012 Jun;263(3):663–72.
2. Newitt DC, Aliu SO, Witcomb N, Sela G, Kornak J, EssermanL, et al. Real-Time Measurement of Functional Tumor Volume by MRI to AssessTreatment Response in Breast Cancer Neoadjuvant Clinical Trials: Validation of theAegis SER Software Platform. Transl Oncol. 2014 Mar;7(1):94–100.