Plataformas Tecnológicas na Vigilância Ativa no Câncer de Próstata: Espectroscopia ATR-FTIR, Análise Metabolômica e Aprendizado de Máquina

Abstract

Active surveillance in clinically localized prostate cancer is an option in approximately 50% of newly diagnosed tumors, but it remains challenging to define which patients require definitive treatment. This study aimed to evaluate two non-invasive, rapid, and cost-effective platforms, metabolomics (LC-MS) and ATR-FTIR, combined with artificial intelligence algorithms to indicate active surveillance with greater precision. Participants diagnosed with clinically localized prostate adenocarcinoma, stratified according to NCCN v. 4 (2024) and submitted to radical prostatectomy, were divided into: Group 1 – Active Surveillance: 12 patients (metabolomics (CLAE/MS Q-TOF ESI) / 8 patients (ATR-FTIR) Group 2 – Definitive Treatment: 7 patients (metabolomics (CLAE/MS Q-TOF ESI) / 3 patients (ATR-FTIR). Serum metabolomic analysis (CL/MS Q-TOF – ESI) was performed using an exploratory strategy: 100% frequency filter, unpaired t-test (p < 0.05) and fold change ≥ 2. Orange v. 3.38.1 software was used with supervised Logistic Regression, Naive Bayes and Stochastic Gradient Descent models. Serum spectra measured by ATR-FTIR, truncated to the lipid (3050–2800 cm⁻¹) and fingerprint (1800–900 cm⁻¹) regions, were processed in Orange 3.3.5 and algorithms were built using supervised Neural Networks, AdaBoost, Random Forest and SVM models. Metabolomic analysis identified 514 metabolites across both groups, of which two were differentially expressed with statistical significance (fold change > 2 and p < 0.05): phosphatidylethanolamine and lysophosphatidylcholine, both elevated in the definitive treatment group, suggesting higher neoplastic activity in this group. Predictive models (Logistic Regression, Naive Bayes and Stochastic Gradient Descent) demonstrated high sensitivity, specificity and accuracy in group separation. In ATR-FTIR, the supervised SVM model achieved 100% sensitivity, specificity and accuracy in separating the groups, and SHAP analysis identified the most contributive spectral peaks: 1423 cm⁻¹ (mixed ring vibrations and CH bending) and 3011 cm⁻¹ (N–H stretching). Exploratory metabolomics (CL-MS) identified two promising markers for distinguishing the active surveillance group from the intervention-indicated group. By combining both platforms (CL-MS and ATR-FTIR) with AI, it was possible to develop predictive algorithms that perform well in separating these groups. ATR-FTIR pointed in the direction of the metabolites identified by metabolomic analysis (LC-MS). Both the markers and the prediction algorithms hold potential to guide active surveillance with greater certainty, pending validation in larger studies.