New Paradigm of Machine Learning (ML) in Personalized Oncology: Data Trimming for Squeezing More Biomarkers From Clinical Datasets

References

Chat With PDF

Cite

Share

Abstract

Personalized medicine has a huge potential of transforming healthcare standards when selection of therapies according to standard guidelines often fails, which can be the case in oncology (1, 2), endocrinology (3, 4), neurology (3), treatment of infectious diseases (5, 6) and hemostatic disorders (7, 8). Nowadays, personalized approach can be based on a solid fundament of big biomedical data obtained for an individual patient, analyzed vs. comparable datasets for other individual cases with known clinical outcome. This can help, for example, developing new criteria for predicting response of a cancer patient to a certain treatment. The analysis of Big Data in oncology can benefit significantly from being empowered by machine learning (ML) techniques (9–13) tailored for solving this “P vs. N” problem. ML is usually defined as the study of algorithmically-built mathematical models that have been fitted for the portion of data called the training dataset, to make predictions for the similarly-obtained and similarly structured data called the test or validation dataset. Major principles of ML have been formulated more than half a century ago and transformed methodology in many areas such as engineering, physics, banking, defense, agriculture, and meteorology (11, 14). Efficiencies of ML-based predictor/classifier models are described by specific quality metrics such as sensitivity (Sn), specificity (Sp), area under ROC curve (AUC), accuracy rate (ACC), Matthews correlation coefficient (MCC), or by p-values from statistical tests distinguishing one class from another (15). However, it was only in the beginning of XXI century when such ML on Big Data became possible in biomedicine, still not having a groundbreaking effect (16). This delay is most probably due to relatively recent emergence of experimental methods generating big biomedical data connected with the sufficiently developed IT infrastructure. Among those game-changing experimental methods the major role was played by next-generation sequencing (NGS) and novel mass-spectrometry approaches which enabled performing whole genome-, transcriptome-, proteome-, andmetabolome analyses relatively fast and cheap (17–19), see Figure 1A. This allowed to feed ML methods with big biomedical data thus generating beneficial outputs, also in the field of clinical medicine. For example, over 150 scientific papers have been indexed in the PubMed repository during last 24 months mentioningmachine learning and drug sensitivity1. Here we will focus on applying ML for personalized medicine, primarily oncology, dealing with attempts to generate as much as possible treatment response biomarkers from mediocre datasets. From the point of view of classical ML approaches, most if not all of the available clinical genetic