Abstract Background: The heterogeneity of TNBC results in mixed responses to NAST: 30-60% of patients (pts) have a pCR to standard chemotherapy with an excellent prognosis. Several methods have been used to predict pCR, most yielding a positive predictive value (PPV) of no greater that 70%. To improve prediction of pCR to NAT, we hypothesized that we can integrate mid-treatment US with molecular profiling to generate a GES, thus reducing the need for escalation of therapy (example: immunotherapy) in select patients. We used data from ARTEMIS, a prospective trial that uses molecular profiling and imaging assessment of TNBC response during therapy to personalize NAT. Methods: Patients begin a planned 4 cycles of AC. Those with substantial volumetric reduction (>=70%) of the primary tumor by US (SVR-US) after AC receive standard taxane-based therapy as the second phase of NAT, while those with resistant disease (including disease progression during AC) are offered therapeutic trials based upon molecular profiling of the pre-treatment biopsy. Pathologic response is assessed at surgical resection. Results: 167 patients had RNAseq, US and pCR data. Overall pCR was 36%. TNBCs with SVR-US after AC (n=101) had significantly higher pCR (55 vs 6%, p<0.001). SVR-US had a PPV of 0.55 and NPV of 0.94 for prediction of pCR. Given the strong NPV, we focused on improving the PPV. In the 101 TNBCs that had SVR-US after AC, we performed differential gene expression comparing those with pCR vs residual disease using 74 TNBCs as a training set and 29 as a validation/test set. Differentially expressed genes (N=500-1000) served as a feature set for a series of machine learning models, including GBM (gradient boosting machines), GLM (generalized linear models), SVM (support vector machines) and CNN (convolutional neural networks) (N train=74). CNN and GLM had similar accuracy, NPV and PPV on the validation set (N=29), therefore GLM was selected as the final model because of ease of interpretability. By combining with SVR-US, we were able to increase the PPV of the tiered model from 0.55 (SVR-US after AC) to 0.89 (SVR-US after AC+GES/GLM) (validation set). Our analysis has validated the predictive value of the GES in patients with SVR-US to 4 cycles of AC, but the entire algorithm (including TNBCs without SVR-US) requires a second validation cohort. However, if the PPV and NPV remain consistent, the impact of this strategy to determine which TNBCs require therapy escalation beyond ACàT is estimated in Table 1. Conclusions: We have created an integrative, tiered model combining two complementary modalities (mid-treatment US assessment of response and GES) that has substantially improved the PPV in assessing pCR to NAST using the ARTEMIS strategy. Table 1: PPV and NPV used to estimate the impact in escalation of therapy*Correct decision: 88% of ptsIncorrect decision: 12% of ptsPredicted pCR=pCR (True positives) Therapy correctly not escalatedPredicted non-pCR=non-pCR (True negatives) Therapy correctly escalatedUnder treatmentOver treatmentPredicted pCR does not=pCR (False positives) Therapy incorrectly not escalatedPredicted non-pCR =pCR (False negatives) Therapy incorrectly escalated34%54%8%4%*example of escalated regimen= taxane + novel agent on clinical trial or taxane + immunotherapy (if FDA approved) Citation Format: Sahil Seth, Gaiane M Rauch, Beatriz Adrada, Helen Piwnica-Worms, Lei Hou, Alastair M Thompson, William F Symmans, Bora Lim, Jason White, Giulio F Draetta, Andrew Futreal, Jeffrey Chang, Stacy Moulder. A tiered algorithm using mid-therapy ultrasound (US) response assessment and a novel gene expression signature (GES) improves the prediction of pathologic complete response (pCR) to neoadjuvant therapy (NAT) in triple-negative breast cancer (TNBC): Results from the ARTEMIS trial (NCT02276443) [abstract]. In: Proceedings of the 2020 San Antonio Breast Cancer Virtual Symposium; 2020 Dec 8-11; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2021;81(4 Suppl):Abstract nr PD9-05.