Spectrochimica acta. Part A, Molecular and biomolecular spectroscopy ·11 Nov 2026 ·Added 13 Nov 2026

Plasma ATR-FTIR spectroscopy combined with machine learning for early classification of response to neoadjuvant endocrine therapy: exploratory results from the ANNE trial.

Alinne Tatiane Faria Silva, Vitor Souza Guimaraes, Izabella da Silva Oliveira, Gustavo Ramos Teixeira, Domicio Carvalho Lacerda, Idam de Oliveira-Junior, Anapaula Hidemi Uema Watanabe, Nilton Onari, Bianca Sakamoto Ribeiro Paiva, Marcelo de Almeida Maia, Yara Cristina de Paiva Maia, Carlos Eduardo Paiva

DOI ↗

Abstract

PURPOSE: Neoadjuvant endocrine therapy (NET) is an effective treatment strategy for luminal breast cancer, but early identification of responders remains challenging. ATR-FTIR spectroscopy combined with machine learning may provide a low-cost, minimally invasive liquid biopsy approach for capturing biochemical information from plasma. METHODS: This exploratory study was embedded within the ANNE trial, a prospective single-arm phase II study evaluating neoadjuvant anastrozole in stage II-III luminal breast cancer. Plasma ATR-FTIR spectra obtained at baseline and after 2-4 weeks of treatment were used to classify response defined by rebiopsy Ki67 and surgical PEPI, respectively. Eight classifier families, including PCA followed by linear discriminant analysis and ridge logistic regression, were evaluated in R. Hyperparameters and the use and configuration of synthetic minority oversampling (SMOTE) were selected jointly using stratified four-fold validation within an outer leave-one-out cross-validation procedure. RESULTS: The Ki67 analysis included 49 patients (41 responders; 8 non-responders), and the PEPI analysis included 36 patients (10 responders; 26 non-responders). The highest observed classifier-specific AUCs were 0.738 for baseline Ki67 classification using an RBF support vector machine and 0.869 for PEPI classification using k-nearest neighbors. The respective accuracies were 0.816 and 0.833, sensitivities were 0.854 and 0.800, specificities were 0.625 and 0.846, and Matthews correlation coefficients were 0.424 and 0.614. Complementary analyses showed that mandatory SMOTE did not improve every classifier. When classifier family was also selected within inner validation, AUCs were 0.616 for Ki67 and 0.854 for PEPI. Spectral features were tentatively associated with lipid-, protein-, nucleic acid- and carbohydrate-related vibrational modes. CONCLUSIONS: Plasma ATR-FTIR spectroscopy combined with machine learning showed exploratory potential for classifying endocrine response, with stronger observed discrimination for PEPI using on-treatment plasma than for Ki67 using baseline plasma. Joint tuning of hyperparameters and balance strategy allowed class-specific performance to be assessed without imposing oversampling on every model. The findings require external validation in larger cohorts before clinical application.