Inverse Analysis of Machine Learning Models for Position Estimation of Biomagnetic Signal Sources

Abstract

Many aspects of how magnetic particles, particularly magnetite particles, are distributed in living organisms and how they affect brain function and neurodegenerative diseases remain unclear. There is an urgent need to develop new methods and techniques to non-invasively and highly accurately detect the presence of these magnetic particles and estimate their location. In this study, we adopted Nearest Neighbors as a machine learning algorithm and analyzed the inverse problem of the machine learning model to estimate the position of magnetic particles that are the source of biomagnetic signals. By arranging the magnetic sensors in three dimensions, the percentage of position estimation errors of 1 cm or less increased, even though there were only 8 magnetic sensors, and the average position estimation error per horizontal plane was approximately 8 mm. Since the resolution of conventional magnetoencephalography equipment is 5 to 7 mm, and measurements are performed using approximately 150 SQUID sensors, it is possible to improve position estimation accuracy by adjusting the placement conditions of the magnetic sensors.