Biomarkers discovery through multivariate statistical methods to face clinical issues concerning thyroid tumour variants classification

Thyroid nodules are common among Western populations, with an estimated prevalence of 50% among individuals aged above 60. However, only 5-10% of the nodules are cancerous, making identifying malignant lesions a substantial health concern for pathologists searching for novel and more accurate diagnostic tools and techniques. Machine Learning (ML) algorithms have emerged as a transformative force in healthcare, improving medical practice in several aspects, including diagnosing tumours. Among the possible biomarkers of thyroid cancer, molecular features obtained through Matrix Assisted Laser Desorption Ionization Mass Spectrometry Imaging (MALDI-MSI) are the most promising. This work presents an application of several ML algorithms using molecular features to build an accurate diagnostic tool for the classification of thyroid nodules [1]. The primary goal of this research is to discover discriminatory molecular signals that can serve as valuable biomarkers. These tumour markers play a crucial role in accurately classifying undefined thyroid cancer variants, such as the Non Invasive Follicular Thyroid Neoplasm with Papillary-like nuclear features type (NIFTP), shedding light on their behaviour and establishing connections to malignancy or benignity. Regarding the ML methods considered for the task, the implementation and comparison of Linear Discriminant Analysis (LDA), Diagonal Discriminant Analysis (DDA), and sparse Partial Least Squares Discriminant Analysis (sPLS-DA) in this work have provided valuable insights into understanding the behaviour of NIFTP. The noteworthy aspect is that all three techniques discover common and relevant features as biomarkers for the NIFTP class, thus improving the reliability of the results from a statistical point of view. These supervised approaches have enabled the identification of specific molecular signals that effectively distinguish thyroid tumour classes, shedding light on NIFTP-type characteristics within this context, achieving accuracy greater than 0.9. This synergy between the medical and machine learning domains can also catalyze further exploration in biomarker discovery. Expanding the applications of supervised learning approaches to address clinical issues in the omics field is a pivotal aspect that can foster cutting-edge research and provide a reliable starting point for researchers to implement and enhance machine learning techniques.