Abstract P2-16-22: Can we improve pCR prediction in early breast cancers treated with neoadjuvant chemotherapy using a simple machine learning algorithm based on 18-FDG and clinico-pathological parameters?

Abstract

Abstract Purpose To test the hypothesis that a simple machine learning algorithm could improve pathological complete response (pCR) prediction in breast cancer patients treated with neoadjuvant chemotherapy (NAC) using an integrative approach based on clinico-pathological and metabolic factors. Methods A total of 311 patients with non-metastatic breast cancer were included. Baseline clinico-pathologic features were evaluated before NAC and tumor uptake of FDG was evaluated before and after one course of NAC. The predictive value of clinico-pathologic and metabolic parameters on pCR was calculated by means of logistic regression analysis and a simple machine learning algorithm generating a hierarchical predictive model using a classification and regression tree (CART). Results: The CART model performed better than all the clinico-pathological and metabolic parameters that were predictive of pCR after univariate analysis. The CART model identified five different groups (G): G1 (luminal cancer with Ki67≤30%), G2 (luminal cancer with Ki67>30% and baseline SUVmax≤10.1), G3 (luminal cancer with Ki67<30% and baseline SUVmax> 10.1), G4 (Her2 positive or triple negative cancer with ΔSUVmax ≤50%) and G5 (Her2 positive or triple negative cancer with ΔSUVmax >50%). The pCR rates predicted by the model for each of the identified groups were G1: 7.2%, G2: 16.7%, G3: 54.6%, G4: 35% and G5: 85% Conclusion This study showed that CART, a simple machine learning algorithm, integrating clinico-pathological and metabolic parameters easily obtainable in routine clinical practice, can improve pCR prediction in early breast cancer and identify subgroups that will not benefit from NAC Baseline patient characteristics of the entire cohort and according to the tumor subtypeCharacteristicsGlobalLumALumBHER2THN311321198476Age (median)4950.249.351.143.3Histological type NST/Other290/2127/5107/1281/375/1Grade 14%22%4%1%0%Grade 237%56%46%32%19%Grade 357%22%47%62%80%Grade missing2%0%3%5%1%ER+58%100%89%45%0%PgR+48%88%75%37%0%ER and PgR neg36%0%0%43%0%HER2 positif27%0%0%100%0%Nodal status negative32%13%32%35%33%Nodal status positive64%50%63%63%61%Nodal status missing data4%37%5%2%6%Stage 11%0%0%0%1%Stage 2A25%28%24%11%28%Stage 2B26%34%24%31%17%Stage 3A16%9%17%13%18%Stage 3B21%22%23%25%17%Stage 3C7%3%7%5%11%Missing stage5%3%5%4%8%mean SUV max (PET-FDG)8.94.859.28.411.7 Univariate and multivariate analyses for pCR predictionVariablepCR (%)OR (95% CI)Univariate analysisp valueOR (95% CI)Multivariate analysisp valueMolecular subtypeLum A1 (3.1)0.15 (0.01-0.076)<0.0010.44 (0.02-2.71)0.029Lum B22 (17.7)11HER241 (48.8)4.42 (2.38-8.41)3.64 (1.40-9.65)TN31 (43.7)3.59 (1.87-7.01)1.65 (0.56-5.02)Histiological grade1/224 (18;8)1<0.001368 (38.6)2.73 (1.61-4.74)Baseline Ki67<=30%26 (19)1<0.001>30%60 (41.4)3.01 (1.77-5.23)AJCC TNM Stage0.36nodal statusN033 (33)0.53N+58 (29)Body surface area<=1.73160 (52.8)2.11 (1.28-3.51)0.003>1.73143 (47.2)1Baseline SUV max<=10.127 (22)1<0.00110.04>10.139 (47.6)3.22 (1.77-5.98)2.24 (1.04-4.67)Delta SUV max<50%17 (22.7)1<0.001>=50%22 (66.7)6.82 (2.83-17.41) Citation Format: Marc-Antoine Benderra, Martine Antoine, Sonia Zilberman, Sandrine Richard, Fatima Kebir, Sofiane Bendifallah, Jean-Pierre Lotz, Emile Darai, Joseph Gligorov, Khaldoun Kerrou. Can we improve pCR prediction in early breast cancers treated with neoadjuvant chemotherapy using a simple machine learning algorithm based on 18-FDG and clinico-pathological parameters? [abstract]. In: Proceedings of the 2019 San Antonio Breast Cancer Symposium; 2019 Dec 10-14; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2020;80(4 Suppl):Abstract nr P2-16-22.