Statistical and Machine Learning Framework for Dynamic-to-Static Mapping: Applications in Neuroscience
Abstract
Statistical and machine learning frameworks are developed for transforming dynamic time series data into static feature representations, with applications in neurophysiological signal analysis. The research utilizes automated feature extraction methods, such as TSFEL, to convert pupil diameter recordings and resting-state functional MRI (rs-fMRI) signals into high-dimensional yet interpretable feature vectors, enabling dimensionality reduction while preserving critical dynamic properties.
For ADHD diagnostics, pupillary time series are mapped into static features across statistical, temporal, and spectral domains. These features are incorporated into supervised classification models, supporting pupillometry as a non-invasive biomarker for neurodevelopmental conditions.
In neuroimaging, rs-fMRI time series from attention-related brain regions undergo similar transformation into static representations. Comparative analyses reveal nonlinear regression models more effectively capture complex brain-behavior relationships than linear models, particularly when predicting ADHD symptoms.
The research emphasizes rigorous model evaluation through cross-validation frameworks optimized for adjusted R² in regression tasks and balanced accuracy in classification tasks. By integrating dynamic-to-static feature mapping, advanced feature selection, and robust modeling strategies, this work presents a reproducible framework for analyzing complex neural signals, contributing objective biomarkers and methodological innovations to statistical learning in neuroscience.