Independent validation of a novel, non-invasive approach to predict pathologic complete response (pCR) in a blinded, prospectively-run single center trial

Background: As pathologic complete response (pCR) is correlated with higher rates of event free survival, accurately forecasting pCR advances our collective endeavors in precision oncology to discern and translate individual patient-specific data into risk stratification. We developed the TumorScope engine, a software platform that utilizes pretreatment diagnostic data to build a computational tumor model that simulates in vivo tumor characteristics and interactions, incorporating morphology, metabolism, vascularity, and nutrient and drug delivery. This non-invasive approach enables accurate forecasting of a patient’s response to physician-chosen neoadjuvant chemotherapy-based treatment (NAC). Here we validate the prognostic capacity of this technology at a single site cancer center. 

Methods: A blinded, prospective trial using retrospective data was conducted at a single institution. The study cohort included patients aged 18 years or older diagnosed with any subtype of breast cancer who were treated with a NAC regimen and had a pre-treatment T1-weighted dynamic contrast enhanced (DCE) MRI available. Pre-treatment diagnostic and planned treatment data (demographics, drug regimen, receptor status (ER/PR/HER2), DCE MRI, and pathology) were input into the TumorScope engine to simulate predicted final tumor volumes (Vt) for each tumor and predict pCR or residual disease (RD); pCR predictions were compared to post-surgery pathologic assessments. Predicted pCR was pre-defined as a residual Vt less than 0.01 cm³, or at least a 99.9% Vt reduction. 

Results: One hundred and fifty subjects with 157 tumors were enrolled in the study. After excluding missing data (absent DCE-MRI), a total of 143 cases in 136 patients were included. The majority of patients self-identified as Caucasian (63%) or African American (23%). TumorScope had a pCR overall prediction accuracy of 92.3% (95% CI: 86.7 – 96.1%) with a sensitivity of 90.9 % (95% CI: 75.7 – 98.1 %) and specificity of 92.7% (95% CI: 86.2 – 96.8%). Based on our subgroup analysis, predictive accuracy remained reliable for HR+/HER2- (n=65; 95.4%), HR+/HER2+ (n=20; 85.0%), HR-/HER2+ (n=21; 85.7%) and TNBC (n=37; 94.6%) subtypes. Predictive performance remained stable across ethnic subtypes and tumor grade (see Table 1). 

Table 1. TumorScope prediction performance.

Conclusion: The TumorScope platform incorporates imaging, pathologic, demographic and planned treatment data appears to accurately predict an individual patient’s probability of pCR across clinical subtypes. These predictions accurately forecast pCR, and the TumorScope technology thus presents itself as a state-of-the-art technology for both physicians and patients during treatment planning.