Model Selection for Magnetic Resonance Spectroscopy - Project Summary/Abstract In-vivo proton magnetic resonance spectroscopy (1H-MRS) can non-invasively measure levels of more than 20 biochemicals in the human brain. With its ability to detect surrogate markers of neuronal health and cell proliferation, neurotransmitters, antioxidants, tumor markers and others, 1H-MRS provides a unique window onto metabolic changes in health and disease. It therefore holds great potential for clinical research, diagnosis, and treatment response monitoring. Clinical translation of MRS has been curbed by the wide range of available technical methods for data acquisition and analysis. These have often produced inconsistent results. Recent research has particularly recognized that metabolite level estimates depend strongly on the way that the measured data are modeled. Traditional 1H-MRS analysis procedures do not capture this uncertainty, and there are currently no methods to determine whether one model is preferable over another. The core theme of this project is a paradigm shift: fitting the measured data with a set of multiple candidate models will replace traditional single-model analysis. A multiverse analysis framework will then capture and quantify the variability across the candidate models, particularly focusing on currently existing MRS software. Model selection tools, in contrast, will use statistical information criteria to actively discriminate which models are the most suitable for a given dataset. These twin strategies (simultaneous consideration of multiple candidate models and data-driven identification of the ‘best’ models) are complementary and provide practical tools to boost the accuracy and precision of metabolite measurement. Clinical and research utility of the new strategies are further amplified by interfacing them with stochastic Markov Chain Monte Carlo sampling. These methods allow model parameter distributions to be characterized more accurately, providing alternative means of uncertainty estimation that eliminate many weaknesses of traditional Cramer-Rao Lower Bounds (CRLB). The project will further demonstrate that these new methods improve the accuracy and precision of non-invasive measurement of 2-hydroxyglutarate (2-HG), a highly specific oncometabolite that plays a pivotal role in IDH- mutated low-grade glioma. In summary, this project proposes several novel data analysis strategies for in-vivo 1H-MRS that address the challenge of analytic variability associated with the choice of modeling approach. All developed software code will be made available to the community through our well-established open-source software ‘Osprey’.
