Abstract P1-08-16: Using machine learning approaches to predict response to neoadjuvant chemotherapy in patients with triple-negative breast cancer

Abstract

Abstract Background: Pathological complete response (pCR) after neoadjuvant chemotherapy (NAC) isassociated with a favorable prognosis and low recurrence rate, particularly in patients with triple-negative breast cancer (TNBC), an aggressive breast cancer subtype. However, only 30%-40%of patients with TNBC show pCR after standard NAC, while the remaining patients have non-pCR or residual disease. Different histological components of the tumor, including tumor stroma,polyploid giant cancer cells, and immune cells, can influence response to NAC in patients withTNBC. However, it remains unknown which histological components contribute to pCR afterNAC in some patients with TNBC but not in others. To address this, we developed a machinelearning pipeline to identify the most significant histological components contributing to NACresponse in patients with TNBC. Different representations of histological components were usedin this pipeline as a "full-slide" feature set to predict NAC response in patients with TNBC.Methods: Post-NAC tissues from 129 patients (55 pCR and 74 non-pCR) from Emory DecaturHospital (Decatur, GA) were used. The Decatur cohort was separated based on NAC responseinto training (33 pCR and 33 non-pCR) and validation sets. There were a total of 17 histologicalclasses annotated by a board-certified pathologist and two NAC labels. Our pipeline comparedfour different machine learning approaches (1NN, ensembleTree, linSVM, and rbfSMV), and thebest classifiers were used for testing. Each whole slide image (WSI) was partitioned into smallertiles. The first classifier used texture-based features from the 17 histological classes. Inaggregation, 55 texture features derived from each tile were provided to a classifier (rbfSVM) tooutput a probability for each new tile during testing. The predicted histology labels were used togenerate a tile-level classification map that recaptured NAC response to predict pCR or non-pCR. Each such tile-level classification map was represented by 80 graph-based featurescapturing the relevant spatial information across the different histological classes. These graph-based features were provided to the best patient-level classifier (rbfSVM) to predict NACresponse for each patient.Results: Our tile-level classification achieved 83% histological classification accuracy, and ourpatient-level machine learning pipeline achieved a 69% NAC classification accuracy. The keyhistological classes that helped distinguish between pCR and non-pCR were tumor & immunecells, tumor & stroma, and microvessel density (MVD) & immune cells. Patients with thesecombinations of discriminating histological classes were 69% likely to have a pCR after NACbased on the accuracy of the rbfSVM NAC classifier. This finding suggests that a classificationmap with more pCR combinations in WSI classified by rbfSVM NAC classifier can helpdistinguish pCR from non-pCR. Notably, our pipeline required less than 24 hours to analyzeWSIs of one patient, accelerating patient selection for NAC treatment.Conclusion: Our machine learning pipeline can robustly identify the most significanthistological classes predicting NAC response and can guide patient selection for NAC treatment. Citation Format: Timothy Byron Fisher, Hongxiao Li, Rekha TS, Jayashree Krishnamurthy, Shristi Bhattarai, Emiel A.M. Janssen, Jun Kong, Ritu Aneja. Using machine learning approaches to predict response to neoadjuvant chemotherapy in patients with triple-negative breast cancer [abstract]. In: Proceedings of the 2021 San Antonio Breast Cancer Symposium; 2021 Dec 7-10; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2022;82(4 Suppl):Abstract nr P1-08-16.